CN111814902A - Target detection model training method, target recognition method, device and medium - Google Patents

Abstract

The embodiment of the application provides a target detection model training method, a target identification device and a target detection model training medium. Will N training images are as the sample set of initial detection model, very big improvement the richness of training sample, make target detection model can detect or discern the target from a plurality of dimensions, can effectively avoid the hourglass of target to examine very big improvement target detection model is to the degree of accuracy of target detection or discernment, and it is unsatisfactory to have solved the detection effect of the detection model that exists among the prior art, takes place easily to miss the technical problem of the condition of target, has reached the technological effect that improves target detection model comprehensiveness and the degree of accuracy.

Description

Translated fromChinese

目标检测模型训练方法、目标识别方法、装置和介质Target detection model training method, target recognition method, device and medium

技术领域technical field

本申请涉及目标检测技术领域，特别是涉及一种目标检测模型训练方法、目标识别方法、装置和介质。The present application relates to the technical field of target detection, and in particular, to a target detection model training method, target recognition method, device and medium.

背景技术Background technique

随着人工智能的发展，可以通过无人机等设备对于目标进行拍摄及识别，以实现对不同环境与不同类别的目标进行快速准确识别与检测，以及时发现并确定目标的位置。在目标识别过程中，深度学习可以代替人工来检测目标的类别、目标各零部件位置、形貌特征等。目前一般通过无人机等对目标进行拍摄，将拍摄到的多张目标图像输入至服务器等处理设备，由嵌有检测模型的处理设备对目标进行深度学习的检测。但是，现有的检测模型的检测效果并不理想，容易发生漏掉目标的情况。With the development of artificial intelligence, the target can be photographed and identified by drones and other equipment, so as to achieve rapid and accurate identification and detection of targets in different environments and different categories, and to find and determine the location of the target in time. In the process of target recognition, deep learning can replace manual work to detect the category of the target, the position of each component of the target, and the topographic features. At present, the target is generally photographed by drones, etc., and the multiple target images captured are input to the processing equipment such as the server, and the processing equipment embedded with the detection model performs deep learning detection on the target. However, the detection effect of the existing detection model is not ideal, and it is easy to miss the target.

发明内容SUMMARY OF THE INVENTION

基于此，有必要针对上述技术问题，提供一种目标检测模型训练方法、目标识别方法、装置和介质。Based on this, it is necessary to provide a target detection model training method, target recognition method, device and medium for the above technical problems.

一种目标检测模型训练方法，所述方法包括：A method for training a target detection model, the method comprising:

对M张样本图像进行数据增广处理，得到N张训练图像，其中M、N均为不小于1的正整数，且M＜N；Perform data augmentation processing on M sample images to obtain N training images, where M and N are both positive integers not less than 1, and M<N;

将所述N张训练图像输入至初始检测模型进行训练，得到第一训练模型；The N training images are input into the initial detection model for training to obtain the first training model;

对所述第一训练模型进行稀疏化和剪枝训练，得到第二训练模型；Perform sparse and pruning training on the first training model to obtain a second training model;

将所述N张训练图像输入至所述第二训练模型进行训练，得到目标检测模型。The N training images are input into the second training model for training to obtain a target detection model.

在其中一个实施例中，所述N张训练图像包括：第一训练图像、第二训练图像和第三训练图像；所述对所述M张样本图像进行数据增广处理，得到N张训练图像，包括：In one embodiment, the N training images include: a first training image, a second training image, and a third training image; and performing data augmentation processing on the M sample images to obtain N training images ,include:

对所述M张样本图像进行切图数据增广处理，得到N1张第一训练图像，其中N1为不小于1的正整数，且M＜N1＜N；Perform image-cut data augmentation processing on the M sample images to obtain N1 first training images, where N1 is a positive integer not less than 1, and M<N1<N;

对所述M张样本图像进行HSV色域变化数据增广处理，得到N2张第二训练图像，其中N2为不小于1的正整数，且M＜N2＜N；Perform HSV color gamut change data augmentation processing on the M sample images to obtain N2 second training images, where N2 is a positive integer not less than 1, and M<N2<N;

对所述M张样本图像进行马赛克数据增广处理，得到N3张第三训练图像，其中N1为不小于1的正整数，且M＜N3＜N。Perform mosaic data augmentation processing on the M sample images to obtain N3 third training images, where N1 is a positive integer not less than 1, and M<N3<N.

在其中一个实施例中，所述N1张第一训练图像包括第一切图图像和第二切图图像；所述对所述M张样本图像进行切图数据增广处理，得到N1张第一训练图像，包括：In one embodiment, the N1 first training images include a first cut image and a second cut image; the cut data augmentation processing is performed on the M sample images to obtain N1 first cut images Training images, including:

对所述M张样本图像进行第一切图处理，得到N11张第一切图图像，其中N11为不小于1的正整数，且M＜N11＜N1；Perform the first cut-through processing on the M sample images to obtain N11 first cut-through images, where N11 is a positive integer not less than 1, and M<N11<N1;

对所述M张样本图像进行第二切图处理，得到N12张第二切图图像，其中N12为不小于1的正整数，且M＜N12＜N1，所述第二切图处理与所述第一切图切图方式不同。Perform second image-cutting processing on the M sample images to obtain N12 second image-cutting images, where N12 is a positive integer not less than 1, and M<N12<N1, and the second image-cutting processing is the same as the The first all maps are cut in different ways.

在其中一个实施例中，所述对所述M张样本图像进行切图数据增广处理，得到N1张第一训练图像，包括：In one of the embodiments, performing image-cutting data augmentation processing on the M sample images to obtain N1 first training images, including:

按照预设模型对所述M张样本图像进行切图处理，得到M1张切图图像；Perform image-slicing processing on the M sample images according to the preset model to obtain M1 sliced images;

对所述M1张切图图像进行筛选，去除无目标图像，得到所述N1张第一训练图像，所述无目标图像是指图像中未出现目标的图像。The M1 sliced images are screened to remove non-target images to obtain the N1 first training images, where the non-target images refer to images in which the target does not appear.

在其中一个实施例中，所述对所述第一训练模型进行稀疏化和剪枝训练，得到第二训练模型，包括：In one embodiment, performing sparse and pruning training on the first training model to obtain a second training model, including:

确定预设区间，所述预设区间是指预设的缩放因子数值范围；determining a preset interval, where the preset interval refers to a preset scaling factor numerical range;

去除所述第一训练模型中所述预设区间内的缩放因子对应的通道，得到所述第二训练模型。The channel corresponding to the scaling factor in the preset interval in the first training model is removed to obtain the second training model.

在其中一个实施例中，所述确定预设区间，包括：In one embodiment, the determining a preset interval includes:

对所述第一训练模型中的多个缩放因子按照大小进行排序，得到缩放因子序列表；Sorting the multiple scaling factors in the first training model according to size to obtain a scaling factor sequence list;

在所述缩放因子序列表中，从最小数值的所述缩放因子开始，将连续的预设数量的所述缩放因子确定为所述预设区间。In the scaling factor sequence list, starting from the scaling factor with the smallest value, a continuous preset number of the scaling factors is determined as the preset interval.

在其中一个实施例中，所述将所述N张训练图像输入至所述第二训练模型进行训练，得到目标检测模型，包括：In one embodiment, the N training images are input into the second training model for training to obtain a target detection model, including:

将所述N张训练图像输入至所述第二训练模型进行训练，得到中间检测模型；Inputting the N training images into the second training model for training to obtain an intermediate detection model;

计算所述中间检测模型的平均精度均值；calculating the mean precision of the intermediate detection model;

若所述平均精度均值不小于预设阈值，则确定所述中间检测模型为所述目标检测模型。If the average precision is not less than a preset threshold, the intermediate detection model is determined to be the target detection model.

在其中一个实施例中，还包括：In one embodiment, it also includes:

若所述平均精度均值小于预设阈值，则将所述中间检测模型确定为所述第一训练模型，并返回执行所述对所述第一训练模型进行稀疏化和剪枝训练，得到第二训练模型。If the average precision is smaller than the preset threshold, the intermediate detection model is determined as the first training model, and the process of performing the sparse and pruning training on the first training model is returned to obtain a second Train the model.

一种目标检测模型训练装置，所述装置包括：A target detection model training device, the device includes:

数据增广模块，用于对M张样本图像进行数据增广处理，得到N张训练图像，其中M、N均为不小于1的正整数，且M＜N；The data augmentation module is used to perform data augmentation processing on M sample images to obtain N training images, where M and N are both positive integers not less than 1, and M<N;

模型训练模块，用于将所述N张训练图像输入至初始检测模型进行训练，得到第一训练模型；对所述第一训练模型进行稀疏化和剪枝训练，得到第二训练模型；将所述N张训练图像输入至所述第二训练模型进行训练，得到目标检测模型。A model training module for inputting the N training images into an initial detection model for training to obtain a first training model; performing sparse and pruning training on the first training model to obtain a second training model; The N training images are input to the second training model for training to obtain a target detection model.

一种目标识别方法，包括：A target recognition method, comprising:

获取待识别图像；Get the image to be recognized;

将所述待识别图像输入至所述目标检测模型，通过所述目标检测模型对所述待识别图像进行目标识别处理，其中，所述目标检测模型通过如上所述的目标检测模型训练方法训练得到；Input the to-be-recognized image into the target detection model, and perform target recognition processing on the to-be-recognized image by the target detection model, wherein the target detection model is obtained by training the above-mentioned target detection model training method ;

根据所述目标检测模型的目标识别处理结果确定目标。The target is determined according to the target recognition processing result of the target detection model.

一种目标识别装置，所述装置包括：A target identification device, the device comprising:

图像获取模块，用于获取待识别图像；an image acquisition module, used to acquire an image to be recognized;

目标识别模块，用于将所述待识别图像输入至所述目标检测模型，通过所述目标检测模型对所述待识别图像进行目标识别处理，其中，所述目标检测模型通过如上所述的目标检测模型训练方法训练得到；A target recognition module, configured to input the to-be-recognized image into the target detection model, and perform target recognition processing on the to-be-recognized image through the target detection model, wherein the target detection model passes the above-mentioned target The detection model training method is trained;

目标确定模块，用于根据所述目标检测模型的目标识别处理结果确定目标。The target determination module is used for determining the target according to the target recognition processing result of the target detection model.

一种计算机设备，包括存储器和处理器，所述存储器存储有计算机程序，所述处理器执行所述计算机程序时实现如上所述的方法的步骤。A computer device includes a memory and a processor, the memory stores a computer program, and the processor implements the steps of the method as described above when the processor executes the computer program.

一种计算机可读存储介质，其上存储有计算机程序，所述计算机程序被处理器执行时实现如上所述的方法的步骤。A computer-readable storage medium having stored thereon a computer program that, when executed by a processor, implements the steps of the method as described above.

本申请实施例提供了一种目标检测模型训练方法，对预先获取的M张样本图像进行数据增广处理，得到N张训练图像，也就是形成训练样本集。然后基于所述训练样本集进行训练和处理得到目标检测模型。本申请实施例通过对所述M张样本图像进行数据增广处理，可以得到更为丰富的训练图像。将所述N张训练图像作为所述初始检测模型的样本集，极大的提高了训练样本的丰富性，使得所述目标检测模型可以从多个维度对目标进行检测或识别，可以有效避免目标的漏检极大的提高所述目标检测模型对于目标检测或者识别的准确度，解决了现有技术中存在的检测模型的检测效果不理想，容易发生漏掉目标的情况的技术问题，达到了提高目标检测模型全面性和准确度的技术效果。The embodiment of the present application provides a target detection model training method, which performs data augmentation processing on M pre-acquired sample images to obtain N training images, that is, a training sample set is formed. Then, the target detection model is obtained by training and processing based on the training sample set. In this embodiment of the present application, more abundant training images can be obtained by performing data augmentation processing on the M sample images. Using the N training images as the sample set of the initial detection model greatly improves the richness of the training samples, so that the target detection model can detect or identify the target from multiple dimensions, which can effectively avoid the target The missed detection greatly improves the accuracy of the target detection model for target detection or recognition, solves the technical problem that the detection effect of the detection model existing in the prior art is not ideal and the target is easily missed, and achieves The technical effect of improving the comprehensiveness and accuracy of the target detection model.

附图说明Description of drawings

图1为一个实施例中目标检测模型训练方法和目标识别方法的应用环境图；Fig. 1 is the application environment diagram of target detection model training method and target recognition method in one embodiment;

图2为一个实施例中目标检测模型训练方法的流程示意图；2 is a schematic flowchart of a method for training a target detection model in one embodiment;

图3为一个实施例中目标检测模型训练方法的流程示意图；3 is a schematic flowchart of a method for training a target detection model in one embodiment;

图4为一个实施例中目标检测模型训练方法的流程示意图；4 is a schematic flowchart of a method for training a target detection model in one embodiment;

图5为一个实施例中目标检测模型训练方法的流程示意图；5 is a schematic flowchart of a method for training a target detection model in one embodiment;

图6为一个实施例中目标检测模型训练方法的流程示意图；6 is a schematic flowchart of a target detection model training method in one embodiment;

图7为一个实施例中目标检测模型训练方法的流程示意图；7 is a schematic flowchart of a method for training a target detection model in one embodiment;

图8为一个实施例中目标检测模型训练方法的流程示意图；8 is a schematic flowchart of a method for training a target detection model in one embodiment;

图9为一个实施例中目标识别方法的流程示意图；9 is a schematic flowchart of a target recognition method in one embodiment;

图10为一个实施例中目标检测模型训练装置的结构框图；10 is a structural block diagram of an apparatus for training a target detection model in one embodiment;

图11为一个实施例中目标识别装置的结构框图。FIG. 11 is a structural block diagram of a target identification device in one embodiment.

具体实施方式Detailed ways

为了使本申请的目的、技术方案及优点更加清楚明白，以下结合附图及实施例，对本申请进行进一步详细说明。应当理解，此处描述的具体实施例仅仅用以解释本申请，并不用于限定本申请。In order to make the purpose, technical solutions and advantages of the present application more clearly understood, the present application will be described in further detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present application, but not to limit the present application.

请参见图1，本申请实施例提供的一种目标检测模型训练方法和目标识别方法可以应用于计算机设备，该计算机设备的内部结构图可以如图1所示。该计算机设备包括通过系统总线连接的处理器、存储器、网络接口和数据库。其中，该计算机设备的处理器用于提供计算和控制能力。该计算机设备的存储器包括非易失性存储介质、内存储器。该非易失性存储介质存储有操作系统、计算机程序和数据库。该内存储器为非易失性存储介质中的操作系统和计算机程序的运行提供环境。该计算机设备的数据库用于存储数据。该计算机设备的网络接口用于与外部的终端通过网络连接通信。该计算机程序被处理器执行时以实现一种目标检测模型训练方法和目标识别方法。Referring to FIG. 1 , a method for training a target detection model and a method for recognizing a target provided by an embodiment of the present application may be applied to a computer device, and an internal structure diagram of the computer device may be shown in FIG. 1 . The computer device includes a processor, memory, a network interface, and a database connected by a system bus. Among them, the processor of the computer device is used to provide computing and control capabilities. The memory of the computer device includes a non-volatile storage medium, an internal memory. The nonvolatile storage medium stores an operating system, a computer program, and a database. The internal memory provides an environment for the execution of the operating system and computer programs in the non-volatile storage medium. The database of the computer device is used to store data. The network interface of the computer device is used to communicate with an external terminal through a network connection. The computer program implements a target detection model training method and target recognition method when executed by the processor.

请参见图2，本申请一个实施例提供了一种目标检测模型训练方法和目标识别方法，可以应用于计算机设备、服务器、处理器等，以下实施例以该方法应用于图1中的计算机设备，用于对目标检测模型训练为例进行说明，包括以下步骤：Referring to FIG. 2, an embodiment of the present application provides a target detection model training method and a target recognition method, which can be applied to computer equipment, servers, processors, etc. The following embodiment applies this method to the computer equipment in FIG. 1. , which is used to illustrate the training of the target detection model as an example, including the following steps:

S100、对M张样本图像进行数据增广处理，得到N张训练图像，其中M、N均为不小于1的正整数，且M＜N。S100. Perform data augmentation processing on M sample images to obtain N training images, where M and N are both positive integers not less than 1, and M<N.

所述样本图像是指包含有目标的图像，所述样本图像作为训练的最原始图像，也就是最初始的图像训练样本集。所述训练图像是经过所述数据增广处理后的训练图像，也就是作为后续处理的样本集。所述数据增广用于增加训练图像数据集的丰富性，让所述训练图像数据集尽可能的多样化，从而使得训练得到的模型具有更强的泛化能力。通过所述数据增广处理对所述样本图像进行多样化处理，从而得到比所述样本图像数量更多的所述训练图像，以增加目标的背景丰富性。所述数据增广的方式可以通过水平/垂直翻转、旋转、缩放、裁剪、剪切、平移、对比度、色彩抖动、噪声等方式进行，本实施例不作具体限定，可根据实际情况具体选择或者设定。The sample image refers to an image containing a target, and the sample image serves as the most original image for training, that is, the most initial image training sample set. The training image is a training image after the data augmentation process, that is, a sample set for subsequent processing. The data augmentation is used to increase the richness of the training image data set, so that the training image data set is as diverse as possible, so that the model obtained by training has a stronger generalization ability. Diversify the sample images through the data augmentation process, so as to obtain more training images than the sample images, so as to increase the background richness of the target. The data augmentation method can be performed by horizontal/vertical flip, rotation, zoom, crop, cut, translation, contrast, color jitter, noise, etc., which is not specifically limited in this embodiment, and can be selected or set according to the actual situation. Certainly.

S200、将所述N张训练图像输入至初始检测模型进行训练，得到第一训练模型。S200. Input the N training images into an initial detection model for training to obtain a first training model.

所述初始检测模型可以采用YOLO V1、YOLO V2或YOLO V3等训练模型，每个所述初始检测模型具有多个BN(Batch Normalization)层，每个所述BN层代表一个通道，具有一个确定的缩放因子，不同的BN层对应具有不同的所述缩放因子。同理，训练得出的所述第一训练模型的每个所述BN层也对应有一个所述缩放因子。将所述N张训练图像输入至所述初始检测模型进行训练，例如：以像素为单位对所述N张训练图像进行数据处理和训练，例如可以进行特征提取、特征降维、特征空值处理、特征转换(one-hot)、特征归一化、目标值空值处理，目标值转换等以得到所述第一训练模型。本实施例中的训练可以建立在pytorch深度学习框架上，例如选用YOVO V3模型作为所述初始检测模型以提高所述训练模型的精度。所述N张训练图像的像素可以为608×608，批次大小为128，使用wramup学习率策略，对所述N张训练图像训练70个epoch。本实施例对于所述初始检测模型进行训练的具体过程不作限定，只需要满足基于所述N张训练图像的样本集训练得出所述第一训练模型的目的即可。The initial detection model can adopt training models such as YOLO V1, YOLO V2 or YOLO V3, each of the initial detection models has a plurality of BN (Batch Normalization) layers, and each of the BN layers represents a channel with a certain Scaling factor, different BN layers have different scaling factors. Similarly, each of the BN layers of the first training model obtained by training also corresponds to one of the scaling factors. Input the N training images into the initial detection model for training, for example: perform data processing and training on the N training images in units of pixels, such as feature extraction, feature dimensionality reduction, and feature null value processing , feature conversion (one-hot), feature normalization, target value null value processing, target value conversion, etc. to obtain the first training model. The training in this embodiment can be established on the pytorch deep learning framework, for example, the YOVO V3 model is selected as the initial detection model to improve the accuracy of the training model. The pixels of the N training images may be 608×608, the batch size is 128, and the N training images are trained for 70 epochs using the wramup learning rate strategy. This embodiment does not limit the specific process of training the initial detection model, and only needs to satisfy the purpose of obtaining the first training model by training based on the sample set of the N training images.

S300、对所述第一训练模型进行稀疏化和剪枝训练，得到第二训练模型。S300. Perform sparse and pruning training on the first training model to obtain a second training model.

所述稀疏化是指对所有BN层的所述缩放因子进行惩罚，减小所述缩放因子的数量，从而减小所述第一训练模型的体积和内部运算量。所述剪枝是指对所述第一训练模型中的部分通道进行去除，例如可以根据实际需要，按照一定规则对所述第一模型的部分通道进行去除。所述第二训练模型是指对所述第一训练模型进行稀疏化和剪枝训练后得到的训练模型，所述稀疏化和剪枝训练可以采用范式函数，例如L1 norm损失函数、L2 norm损失函数等，本实施例对于所述稀疏化和剪枝训练的具体方式不作限定，可根据实际情况具体选择或者设定，只需要满足可以实现对所述第一训练模型进行稀疏化和剪枝训练的功能即可。The sparsification refers to penalizing the scaling factors of all BN layers to reduce the number of scaling factors, thereby reducing the volume of the first training model and the amount of internal computation. The pruning refers to removing part of the channels in the first training model. For example, according to actual needs, part of the channels in the first model can be removed according to certain rules. The second training model refers to a training model obtained by performing sparse and pruning training on the first training model, and the sparse and pruning training may use a normal form function, such as L1 norm loss function, L2 norm loss This embodiment does not limit the specific method of the sparse and pruning training, which can be selected or set according to the actual situation, and only needs to meet the requirements that the first training model can be sparsely trained and pruned. function.

S400、将所述N张训练图像输入至所述第二训练模型进行训练，得到目标检测模型。S400. Input the N training images into the second training model for training to obtain a target detection model.

所述N张训练图像是经过步骤S100对原始样本图像进行数据增广处理过的图像，所述第二训练模型是经过步骤S300进行稀疏化和剪枝训练得到训练模型，将所述N张训练图像作为所述第二训练模型的样本集对所述第二训练模型进行训练，训练方法可以与步骤S200中对所述第一训练模型的训练方法相同，例如均可以使用wramup学习率策略，对所述N张训练图像训练例如100个epoch，以得到所述目标检测模型。本实施例对于所述第二训练模型进行训练的具体过程不作限定，只需要满足基于所述N张训练图像的样本集训练得出所述目标检测模型的目的即可。The N training images are images obtained by performing data augmentation processing on the original sample images in step S100, and the second training model is a training model obtained by performing sparse and pruning training in step S300, and the N training images are The image is used as the sample set of the second training model to train the second training model, and the training method can be the same as the training method of the first training model in step S200, for example, the wramup learning rate strategy can be used to The N training images are trained for, for example, 100 epochs to obtain the target detection model. This embodiment does not limit the specific process of training the second training model, and only needs to meet the purpose of obtaining the target detection model by training based on the sample set of the N training images.

本申请实施例提供了一种目标检测模型训练方法，对预先获取的M张样本图像进行数据增广处理，得到N张训练图像，也就是形成训练样本集。然后基于所述训练样本集进行训练和处理得到目标检测模型。本申请实施例通过对所述M张样本图像进行数据增广处理，可以得到更为丰富的训练图像。将所述N张训练图像作为所述初始检测模型的样本集，极大的提高了训练样本的丰富性，使得所述目标检测模型可以从多个维度对目标进行检测或识别，可以有效避免目标的漏检极大的提高所述目标检测模型对于目标检测或者识别的准确度，解决了现有技术中存在的检测模型的检测效果不理想，容易发生漏掉目标的情况的技术问题，达到了提高目标检测模型全面性和准确度的技术效果。The embodiment of the present application provides a target detection model training method, which performs data augmentation processing on M pre-acquired sample images to obtain N training images, that is, a training sample set is formed. Then, the target detection model is obtained by training and processing based on the training sample set. In this embodiment of the present application, more abundant training images can be obtained by performing data augmentation processing on the M sample images. Using the N training images as the sample set of the initial detection model greatly improves the richness of the training samples, so that the target detection model can detect or identify the target from multiple dimensions, which can effectively avoid the target The missed detection greatly improves the accuracy of the target detection model for target detection or recognition, solves the technical problem that the detection effect of the detection model existing in the prior art is not ideal and the target is easily missed, and achieves The technical effect of improving the comprehensiveness and accuracy of the target detection model.

请参见图3，在一个实施例中，所述N张训练图像包括：第一训练图像、第二训练图像和第三训练图像。所述步骤S100包括：步骤S110-步骤S130。Referring to FIG. 3, in one embodiment, the N training images include: a first training image, a second training image, and a third training image. The step S100 includes: step S110-step S130.

S110、对所述M张样本图像进行切图数据增广处理，得到N1张第一训练图像，其中N1为不小于1的正整数，且M＜N1＜N。S110. Perform image-cut data augmentation processing on the M sample images to obtain N1 first training images, where N1 is a positive integer not less than 1, and M<N1<N.

所述切图数据增广处理是指按照一定的规则对每张所述样本图像进行切图处理，例如可以对每张所述样本图像按照上下左右不同区域进行切割，或者以中心点为圆心，按一定的角度进行扇形切割，或者采用其他方式或者随机切割等均可。本实施例对于切割规则不作具体限定，可根据实际情况具体设定，只需要满足可以实现通过对所述样本图像进行切割以得到更多的所述第一训练图像的功能即可。例如当所述样本图像的数量为1万张时，经所述切图数据增广处理后的图像大约为3万张，即所述第一训练图像的数量N1为3万。目标所处的场景差异性较大，通过切图数据增广可以有效增加各种场景的负样本，将经过所述切图数据增广处理后的所有的包括无目标和有目标的图像集合起来，以减少在后续检测过程中可能出现的误识别或者漏检，从而提高对于本实施例所述目标检测模型目标识别的准确性。The image-slicing data augmentation processing refers to performing image-slicing processing on each of the sample images according to certain rules. For example, each sample image can be sliced according to different regions of the upper, lower, left, and right, or the center point is the center of the circle. Sector cutting at a certain angle, or other methods or random cutting can be used. The cutting rules are not specifically limited in this embodiment, and can be specifically set according to actual conditions, as long as the function of obtaining more first training images by cutting the sample images is satisfied. For example, when the number of the sample images is 10,000, the number of images after the image-cut data augmentation processing is about 30,000, that is, the number N1 of the first training images is 30,000. The scene where the target is located is quite different, and the negative samples of various scenes can be effectively increased by the image-cut data augmentation, and all the images including non-target and target-targeted images after the image-cut data augmentation processing are collected. , so as to reduce possible misrecognition or missed detection in the subsequent detection process, thereby improving the accuracy of target recognition for the target detection model described in this embodiment.

S120、对所述M张样本图像进行HSV色域变化数据增广处理，得到N2张第二训练图像，其中N2为不小于1的正整数，且M＜N2＜N。S120. Perform HSV color gamut change data augmentation processing on the M sample images to obtain N2 second training images, where N2 is a positive integer not less than 1, and M<N2<N.

所述HSV色域是指针对图像的色调H、饱和度S和明度V。色调H是指所述样本图像的颜色，饱和度S是指所述样本图像的深浅，明度V是指所述样本图像的明暗变化。所述HSV色域变化数据增广是指通过对所述样本图像的色调H、饱和度S和明度V的参数进行调整，以增加具有不同色调H、饱和度S和明度V参数的图像，从而得到更多的图像，也就是得到更多的所述第二训练图像。例如当所述样本图像的数量M为1万张时，通过所述HSV色域变化数据增广处理后得到的所述第二训练图像的数量N2大约为3万张。本实施例对于所述HSV色域变化数据增广处理的具体过程和所述色调H、饱和度S和明度V参数改变等均不作任何限定，可根据实际情况具体选择或者设定。The HSV color gamut refers to the hue H, saturation S and lightness V of the image. The hue H refers to the color of the sample image, the saturation S refers to the depth of the sample image, and the lightness V refers to the change of light and dark of the sample image. The HSV color gamut change data augmentation refers to adjusting the parameters of hue H, saturation S and lightness V of the sample image to increase images with different parameters of hue H, saturation S and lightness V, thereby Getting more images means getting more of the second training images. For example, when the number M of the sample images is 10,000, the number N2 of the second training images obtained after the HSV color gamut change data augmentation processing is about 30,000. This embodiment does not limit the specific process of the HSV color gamut change data augmentation processing and the changes of the hue H, saturation S, and lightness V parameters, etc., and can be selected or set according to the actual situation.

S130、对所述M张样本图像进行马赛克数据增广处理，得到N3张第三训练图像，其中N1为不小于1的正整数，且M＜N3＜N。S130. Perform mosaic data augmentation processing on the M sample images to obtain N3 third training images, where N1 is a positive integer not less than 1, and M<N3<N.

所述马赛克数据增广是指mosaic数据增广或者mosaic数据增强，通过所述mosaic数据增广例如约一万张图像，也就是在所述M张样本图像的基础上再通过所述马赛克数据增广增加例如1万张样本图像，以丰富目标的背景。所述mosaic数据增广可以针对遮挡问题，通过随机擦除目标的特征模拟遮挡的效果，提高模型的泛化能力，使模型在训练过程中仅通过局部特征便可实现对目标的识别，强化模型对于目标局部特征的认知，弱化模型对于目标全部特征的依赖，同时通过这样的数据进行训练，模型也会对噪声和遮挡更具鲁棒性。The mosaic data augmentation refers to mosaic data augmentation or mosaic data augmentation. For example, about ten thousand images are augmented by the mosaic data augmentation, that is, the mosaic data augmentation is performed on the basis of the M sample images. Widely increase, for example, 10,000 sample images to enrich the background of the target. The mosaic data augmentation can address the occlusion problem, simulate the occlusion effect by randomly erasing the features of the target, and improve the generalization ability of the model, so that the model can recognize the target only through local features in the training process, and strengthen the model. For the cognition of the local features of the target, the dependence of the model on all the features of the target is weakened, and at the same time, the model is more robust to noise and occlusion by training on such data.

请参见图4，在一个实施例中，所述N1张第一训练图像包括第一切图图像和第二切图图像。所述步骤S110包括步骤S111-S112。Referring to FIG. 4 , in one embodiment, the N1 first training images include a first slice image and a second slice image. The step S110 includes steps S111-S112.

S111、对所述M张样本图像进行第一切图处理，得到N11张第一切图图像，其中N11为不小于1的正整数，且M＜N11＜N1。S111. Perform a first cut-through process on the M sample images to obtain N11 first cut-through images, where N11 is a positive integer not less than 1, and M<N11<N1.

所述第一切图处理是指分别对每张所述样本图像按照一定的规则分别截图生成新的图像，以形成更多的训练图像，也就是得到N11张第一切图图像，将所述N11张第一切图图像增加至原始的M张所述样本图像中，以丰富所述样本图像集。所述第一切图处理可以为按照上左、上中、上右、下左、下中、下右六个区域进行截图，也可以采用其他的规则进行截图，以使得在目标不够完整的情况下，仅通过部分特征便可实现对目标的识别，强化模型对于目标局部特征的认知以及对特定背景的认知，弱化模型对于目标全部特征的依赖。本实施例对于所述第一切图的规则不作具体限定，可根据实际情况具体设定，只需要满足可以得到更多的所述第一切图图像的目的即可。The first image processing refers to taking screenshots of each of the sample images according to certain rules to generate new images to form more training images, that is, to obtain N11 first image images, N11 first cut images are added to the original M sample images to enrich the sample image set. The first all-image processing can be to take screenshots according to six areas of upper left, upper middle, upper right, lower left, lower middle, and lower right, or other rules can be used to take screenshots, so that the target is not complete enough. In this way, the recognition of the target can be realized only through some features, which strengthens the model's cognition of the local features of the target and the cognition of the specific background, and weakens the model's dependence on all the features of the target. This embodiment does not specifically limit the rules of the first drawing, which can be specifically set according to the actual situation, and only needs to meet the purpose of obtaining more images of the first drawing.

S112、对所述M张样本图像进行第二切图处理，得到N12张第二切图图像，其中N12为不小于1的正整数，且M＜N12＜N1，所述第二切图处理与所述第一切图切图方式不同。S112. Perform second image-slicing processing on the M sample images to obtain N12 second image-slicing images, where N12 is a positive integer not less than 1, and M<N12<N1, and the second image-slicing processing is the same as The first cutting diagram is cut in different ways.

所述第二切图处理是指分别对每张所述样本图像按照一定的规则分别截图生成新的图像，以形成更多的训练图像，也就是得到N12张第二切图图像，将所述N12张第二切图图像增加至原始的M张所述样本图像中，以丰富所述样本图像集。所述第二切图处理可以为按照左右两个区域进行截图，也可以采用其他的规则进行截图，但需要指出的是，所述第二切除处理的规则和所述第一切图处理的规则不同。通过所述第二切图处理可以使得所述目标检测模型在目标不够完整的情况下，仅通过部分特征便可实现对目标的识别，强化模型对于目标局部特征的认知以及对特定背景的认知，弱化模型对于目标全部特征的依赖。本实施例对于所述第二切图的规则不作具体限定，可根据实际情况具体设定，只需要满足可以得到更多的所述第二切图图像的目的即可。The second image-cutting process refers to taking screenshots of each of the sample images to generate new images according to certain rules, so as to form more training images, that is, to obtain N12 second-cut images. N12 second cut images are added to the original M sample images to enrich the sample image set. The second cutting process may be to take screenshots according to the left and right regions, or other rules may be used to take screenshots, but it should be pointed out that the rules for the second cutting processing and the rules for the first cutting processing are different. Through the second image-slicing process, the target detection model can recognize the target only through partial features when the target is not complete, and strengthen the model's recognition of the local features of the target and recognition of the specific background. Knowing that, the dependence of the model on all the features of the target is weakened. This embodiment does not specifically limit the rules for the second image cutting, which can be specifically set according to the actual situation, and only needs to meet the purpose of obtaining more second image cutting images.

请参见图5，在一个实施例中，所述S110包括步骤S140-150：Referring to FIG. 5, in one embodiment, the S110 includes steps S140-150:

S140、按照预设模型对所述M张样本图像进行切图处理，得到M1张切图图像。S140. Perform image-slicing processing on the M sample images according to a preset model to obtain M1 sliced images.

所述预设模型是指对于所述M张样本图像进行切图的规则，例如可以按照左右两个区域对每张所述M张样本图像进行截图，或者按照上左、上中、上右、下左、下中、下右六个区域进行截图，或者也可以采用其他截图规则进行截图，以提高所述样本图像的丰富性。本实施例对于所述切图处理的规则或方式等均不作具体限定，可根据实际情况具体选择或者设定。The preset model refers to the rules for cutting the M sample images, for example, each of the M sample images can be screenshots according to the left and right regions, or according to the upper left, upper middle, upper right, The bottom left, bottom middle, and bottom right six regions are taken for screenshots, or other screenshot rules can also be used to take screenshots, so as to improve the richness of the sample image. This embodiment does not specifically limit the rules or manners of the image cutting processing, and may be selected or set according to actual conditions.

S150、对所述M1张切图图像进行筛选，去除无目标图像，得到所述N1张第一训练图像，所述无目标图像是指图像中未出现目标的图像。S150. Screen the M1 cut-out images, remove non-target images, and obtain the N1 first training images, where the non-target images refer to images in which the target does not appear.

所述M1张切图图像是经过切图处理后得到的大量的包含有目标和无目标的图像，例如当所述样本图像的数量M为1万张时，则所述切图图像的数量M1大约为3万张，然后通过筛选去除无目标图像。去除仅包含纯背景的图像，得到剩余的含有目标的图像，也就是所述第一训练图像，以精简所述样本图像，从而在保证所述目标检测模型的准确性的前提下减小所述目标检测模型的体积。The M1 sliced images are a large number of images containing targets and non-targets obtained after image slice processing. For example, when the number M of the sample images is 10,000, the number of sliced images M1 About 30,000 images, and then filtered to remove untargeted images. Remove the images that only contain the pure background, and obtain the remaining images containing the target, that is, the first training image, so as to simplify the sample images, thereby reducing the accuracy of the target detection model under the premise of ensuring The volume of the object detection model.

请参见图6，在一个实施例中，步骤S300包括S310-S320：Referring to FIG. 6, in one embodiment, step S300 includes S310-S320:

S310、确定预设区间，所述预设区间是指预设的缩放因子数值范围。S310. Determine a preset interval, where the preset interval refers to a preset value range of the scaling factor.

每个BN层，也就是每个通道对应有一个缩放因子，所述第一训练模型中具有多个通道，也就是对应有多个所述缩放因子，多个所述缩放因子具有不同的数值。所述预设区间就是根据实际情况对多个所述缩放因子进行取舍的范围，例如需要去除10％～50％之间的所述缩放因子，则所述预设区间即为10％～50％。Each BN layer, that is, each channel corresponds to a scaling factor, and the first training model has multiple channels, that is, corresponds to a plurality of the scaling factors, and the multiple scaling factors have different values. The preset interval is a range in which multiple scaling factors are selected according to actual conditions. For example, if the scaling factors between 10% and 50% need to be removed, the preset interval is 10% to 50%. .

S320、去除所述第一训练模型中所述预设区间内的缩放因子对应的通道，得到所述第二训练模型。S320. Remove the channel corresponding to the scaling factor in the preset interval in the first training model to obtain the second training model.

所述预设区间也就是在步骤S310中确定的按照实际需要确定出的需要去除的所述缩放因子。每个通道也就是每个BN层对应有一个缩放因子，每个所述缩放因子与每个通道都是一一对应的。所述缩放因子与每个通道一一对应，按照所述预设区间对所述第一训练模型进行剪枝处理，也就是去除所述第一模型中所述缩放因子处于所述预设区间内的通道，以得到所述第二训练模型。通过剪枝处理便可实现在保证所述目标检测模型的准确性的前提下减小所述目标检测模型的体积。The preset interval is the scaling factor determined in step S310 and determined according to actual needs and needs to be removed. Each channel, that is, each BN layer corresponds to a scaling factor, and each scaling factor is in one-to-one correspondence with each channel. The scaling factor is in one-to-one correspondence with each channel, and the first training model is pruned according to the preset interval, that is, the scaling factor in the first model is removed from the preset interval. channel to get the second trained model. Through the pruning process, the volume of the target detection model can be reduced on the premise of ensuring the accuracy of the target detection model.

请参见图7，在一个实施例中，步骤S310包括步骤S311-S312：Referring to FIG. 7, in one embodiment, step S310 includes steps S311-S312:

S311、对所述第一训练模型中的多个缩放因子按照大小进行排序，得到缩放因子序列表。S311. Sort the multiple scaling factors in the first training model according to size to obtain a scaling factor sequence list.

所述缩放因子序列表是指将多个所述缩放因子按照大小进行排序形成的缩放因子的排列表。每个通道对应有一个所述缩放因子，则所述第一训练模型中对应有多个所述缩放因子，对所述多个缩放因子按照数值大小进行排序，可以按照由小及大的顺序进行排序，也可以按照由大及小的顺序进行排序，以得到所述缩放因子序列表。The scaling factor sequence table refers to a scaling factor arrangement table formed by sorting a plurality of the scaling factors according to size. Each channel corresponds to one of the scaling factors, then the first training model corresponds to a plurality of the scaling factors, and the multiple scaling factors are sorted according to their numerical values, which can be performed in ascending order. Sorting can also be carried out according to the order of the largest and the smallest, so as to obtain the scaling factor sequence list.

S312、在所述缩放因子序列表中，从最小数值的所述缩放因子开始，将连续的预设数量的所述缩放因子确定为所述预设区间。S312. In the scaling factor sequence table, starting from the scaling factor with the smallest value, determine a continuous preset number of the scaling factors as the preset interval.

所述预设数量是步骤S310中所述预设区间内所述缩放因子的数量，所述预设数量可以按照对所述第一训练模型进行稀疏化程度进行设置，例如可以100、200，或者可以为所述缩放因子总数的80％、90％等。The preset number is the number of the scaling factors in the preset interval in step S310, and the preset number may be set according to the degree of sparseness of the first training model, for example, 100, 200, or It can be 80%, 90%, etc. of the total number of scaling factors.

请参见图8，在一个实施例中，步骤S400包括步骤S410-S440：Referring to FIG. 8, in one embodiment, step S400 includes steps S410-S440:

S410、将所述N张训练图像输入至所述第二训练模型进行训练，得到中间检测模型。S410. Input the N training images into the second training model for training to obtain an intermediate detection model.

所述第二训练模型是对所述第一训练模型进行稀疏化和剪枝训练得到的，本实施例中所述第二训练模型的训练方法和训练过程与步骤S200中对于所述初始检测模型的训练方法可以相同，例如均可以使用wramup学习率策略，对所述N张训练图像训练例如100或200个epoch。本实施例对于所述第二训练模型进行训练的具体过程不作限定，只需要满足基于所述N张训练图像的样本集训练得出所述中间检测模型的目的即可。The second training model is obtained by performing sparse and pruning training on the first training model. The training method and training process of the second training model in this embodiment are the same as those of the initial detection model in step S200. The training methods can be the same, for example, the wramup learning rate strategy can be used to train the N training images for, for example, 100 or 200 epochs. This embodiment does not limit the specific process of training the second training model, and only needs to meet the purpose of obtaining the intermediate detection model by training based on the sample set of the N training images.

S420、计算所述中间检测模型的平均精度均值。S420. Calculate the average precision of the intermediate detection model.

AP(Average precision)值是指平均精度值，将曲线下的面积当做衡量尺度对训练模型，也就是本实施例中的所述中间检测模型的准确率计算平均值。对于模型训练肯定会有多个epoch，对于多次epoch的所述中间检测模型的准确率取平均值，也就得到所述平均精度均值mAP(Mean average precision)。本实施例对于所述平均精度均值的确定方式不作具体限定，可根据实际情况具体选择。The AP (Average precision) value refers to the average precision value, and the area under the curve is used as a measure to calculate the average value of the accuracy of the training model, that is, the intermediate detection model in this embodiment. There must be multiple epochs for model training, and the accuracy of the intermediate detection model for multiple epochs is averaged to obtain the mean average precision mAP (Mean average precision). This embodiment does not specifically limit the manner of determining the mean value of the average precision, and may be selected according to actual conditions.

S430、若所述平均精度均值不小于预设阈值，则确定所述中间检测模型为所述目标检测模型。S430. If the average precision is not less than a preset threshold, determine that the intermediate detection model is the target detection model.

所述预设阈值可以根据实际情况具体选择或者设定，例如可以为50％、60％等。例如所述预设阈值为50％，在对所述中间检测模型训练100个epoch后，对所述中间检测模型进行测试得到的所述中间检测模型的平均精度均值为68.2％，则证明所述中间检测模型的准确度达到预设目标，则停止对于所述中间检测模型的训练，确定当前的所述中间检测模型为所述目标检测模型。The preset threshold may be specifically selected or set according to the actual situation, for example, it may be 50%, 60%, and the like. For example, the preset threshold is 50%. After training the intermediate detection model for 100 epochs, the average precision of the intermediate detection model obtained by testing the intermediate detection model is 68.2%, which proves that the When the accuracy of the intermediate detection model reaches the preset target, the training of the intermediate detection model is stopped, and the current intermediate detection model is determined as the target detection model.

若所述平均精度均值小于预设阈值，则将所述中间检测模型确定为所述第一训练模型，并返回执行步骤S300所述对所述第一训练模型进行稀疏化和剪枝训练，得到第二训练模型。If the average precision value is less than the preset threshold, then determine the intermediate detection model as the first training model, and return to performing the sparse and pruning training on the first training model described in step S300, to obtain The second training model.

对所述中间检测模型训练一定周期后，对所述中间检测模型的平均精度均值进行测试，当所述中间检测模型的平均精度还未达到所述预设阈值时，则证明此时的所述中间检测模型的准确度还比较低。此时，可以返回步骤S300，将所述中间检测模型作为所述第一训练模型，继续执行步骤S300，对所述中间检测模型的稀疏化程度和剪枝通道进一步调整，并继续进行多个周期的训练，以提高所述中间检测模型的准确性。After the intermediate detection model is trained for a certain period, the average precision of the intermediate detection model is tested, and when the average precision of the intermediate detection model has not reached the preset threshold, it is proved that the The accuracy of the intermediate detection model is still relatively low. At this time, it is possible to return to step S300, use the intermediate detection model as the first training model, continue to perform step S300, further adjust the sparsification degree and pruning channel of the intermediate detection model, and continue to perform multiple cycles training to improve the accuracy of the intermediate detection model.

请参见图9，本申请一个实施例提供了一种目标识别方法，包括步骤S500-S700：Referring to FIG. 9, an embodiment of the present application provides a target recognition method, including steps S500-S700:

S500、获取待识别图像。S500. Acquire an image to be recognized.

所述待识别图像可以为任意包括有目标的图像的图像集，所述待识别图像可以为任意场景或者环境下的图像。本实施例对于所述待识别图像的数量或者类别等均不作具体限定，可根据实际情况具体选择。The to-be-recognized image may be any image set including an image with a target, and the to-be-recognized image may be an image in any scene or environment. This embodiment does not specifically limit the number or types of the images to be recognized, and can be selected according to actual conditions.

S600、将所述待识别图像输入至所述目标检测模型，通过所述目标检测模型对所述待识别图像进行目标识别处理，其中，所述目标检测模型通过如上所述的目标检测模型训练方法训练得到。S600. Input the to-be-recognized image into the target detection model, and perform target recognition processing on the to-be-recognized image through the target detection model, wherein the target detection model adopts the above-mentioned target detection model training method Trained to get.

可以先对所述待识别图像的目标识别处理包括对所述待识别图像的预处理，将每张所述待识别图像通过拉伸、压缩等方法统一大小，或者提取每个候选区域的固定长度的特征，然后使用特定类别的线性SVM分类器对每个候选区域进行分类，最近进行BoundingBox回归，会得到一个回归值。然后对所述朳检测模型进行预训练、特征领域的微调、以及将所有候选区域与真实框重叠大于和等于0.5的作为该框类的正例，其余的在进行AVM分类。The target recognition processing of the to-be-recognized image may first include preprocessing of the to-be-recognized image, unifying the size of each of the to-be-recognized images by stretching, compressing, etc., or extracting a fixed length of each candidate area. The features of , then use a class-specific linear SVM classifier to classify each candidate region, and most recently perform a BoundingBox regression, which will get a regression value. Then the pre-training of the detection model, the fine-tuning of the feature field, and the overlap of all the candidate regions with the real box is greater than or equal to 0.5 as the positive example of the box class, and the rest are in AVM classification.

S700、根据所述目标检测模型的目标识别处理结果确定目标。S700. Determine a target according to the target recognition processing result of the target detection model.

通过步骤S600确定所述目标检测模型对于所述待识别图像的识别结果进行预测，根据在步骤S600中得到的回归值与所要预测的Grounding Truth之间的关系，反向推到Grounding Truth的位置或者根据得到的置信度参数等预测目标的位置，以达到确定目标的目的。It is determined by step S600 that the target detection model predicts the recognition result of the image to be recognized, and according to the relationship between the regression value obtained in step S600 and the Grounding Truth to be predicted, reversely push to the position of the Grounding Truth or The position of the target is predicted according to the obtained confidence parameters, etc., so as to achieve the purpose of determining the target.

应该理解的是，虽然流程图中的各个步骤按照箭头的指示依次显示，但是这些步骤并不是必然按照箭头指示的顺序依次执行。除非本文中有明确的说明，这些步骤的执行并没有严格的顺序限制，这些步骤可以以其它的顺序执行。而且，图中的至少一部分步骤可以包括多个步骤或者多个阶段，这些步骤或者阶段并不必然是在同一时刻执行完成，而是可以在不同的时刻执行，这些步骤或者阶段的执行顺序也不必然是依次进行，而是可以与其它步骤或者其它步骤中的步骤或者阶段的至少一部分轮流或者交替地执行。It should be understood that although the various steps in the flowchart are displayed in sequence according to the arrows, these steps are not necessarily executed in sequence according to the arrows. Unless explicitly stated herein, the execution of these steps is not strictly limited to the order, and these steps may be performed in other orders. Moreover, at least a part of the steps in the figure may include multiple steps or multiple stages, and these steps or stages are not necessarily executed at the same time, but may be executed at different times, and the execution order of these steps or stages is also different. It is necessarily performed sequentially, but may be performed alternately or alternately with other steps or at least a portion of the steps or stages within the other steps.

请参见图10，本申请一个实施例提供了一种目标检测模型训练装置10，所述装置包括：数据增广模块100和模型训练模块200。Referring to FIG. 10 , an embodiment of the present application provides a target detectionmodel training apparatus 10 . The apparatus includes: adata augmentation module 100 and amodel training module 200 .

所述数据增广模块100用于对M张样本图像进行数据增广处理，得到N张训练图像，其中M、N均为不小于1的正整数，且M＜N；Thedata augmentation module 100 is configured to perform data augmentation processing on M sample images to obtain N training images, wherein M and N are both positive integers not less than 1, and M<N;

所述模型训练模块200用于将所述N张训练图像输入至初始检测模型进行训练，得到第一训练模型；对所述第一训练模型进行稀疏化和剪枝训练，得到第二训练模型；将所述N张训练图像输入至所述第二训练模型进行训练，得到目标检测模型。Themodel training module 200 is used for inputting the N training images into an initial detection model for training to obtain a first training model; performing sparse and pruning training on the first training model to obtain a second training model; The N training images are input into the second training model for training to obtain a target detection model.

所述数据增广模块100还用于对所述M张样本图像进行切图数据增广处理，得到N1张第一训练图像，其中N1为不小于1的正整数，且M＜N1＜N；对所述M张样本图像进行HSV色域变化数据增广处理，得到N2张第二训练图像，其中N2为不小于1的正整数，且M＜N2＜N；对所述M张样本图像进行马赛克数据增广处理，得到N3张第三训练图像，其中N1为不小于1的正整数，且M＜N3＜N。Thedata augmentation module 100 is further configured to perform image-cut data augmentation processing on the M sample images to obtain N1 first training images, where N1 is a positive integer not less than 1, and M<N1<N; Perform HSV color gamut change data augmentation processing on the M sample images to obtain N2 second training images, where N2 is a positive integer not less than 1, and M<N2<N; The mosaic data augmentation process obtains N3 third training images, wherein N1 is a positive integer not less than 1, and M<N3<N.

所述数据增广模块100还用于对所述M张样本图像进行第一切图处理，得到N11张第一切图图像，其中N11为不小于1的正整数，且M＜N11＜N1；对所述M张样本图像进行第二切图处理，得到N12张第二切图图像，其中N12为不小于1的正整数，且M＜N12＜N1，所述第二切图处理与所述第一切图切图方式不同。Thedata augmentation module 100 is further configured to perform the first cut-through processing on the M sample images to obtain N11 first cut-through images, where N11 is a positive integer not less than 1, and M<N11<N1; Perform second image-cutting processing on the M sample images to obtain N12 second image-cutting images, where N12 is a positive integer not less than 1, and M<N12<N1, and the second image-cutting processing is the same as the The first all maps are cut in different ways.

所述数据增广模块100还用于按照预设模型对所述M张样本图像进行切图处理，得到M1张切图图像；对所述M1张切图图像进行筛选，去除无目标图像，得到所述N1张第一训练图像，所述无目标图像是指图像中未出现目标的图像。Thedata augmentation module 100 is further configured to perform image-slicing processing on the M sample images according to the preset model to obtain M1 sliced images; screen the M1 sliced images to remove non-target images to obtain For the N1 first training images, the non-target images refer to images in which the target does not appear.

所述模型训练模块200还用于确定预设区间，所述预设区间是指预设的缩放因子数值范围；去除所述第一训练模型中所述预设区间内的缩放因子对应的通道，得到所述第二训练模型。Themodel training module 200 is further configured to determine a preset interval, where the preset interval refers to a preset scaling factor numerical range; remove the channel corresponding to the scaling factor in the preset interval in the first training model, Obtain the second training model.

所述模型训练模块200还用于对所述第一训练模型中的多个缩放因子按照大小进行排序，得到缩放因子序列表；在所述缩放因子序列表中，从最小数值的所述缩放因子开始，将连续的预设数量的所述缩放因子确定为所述预设区间。Themodel training module 200 is further configured to sort a plurality of scaling factors in the first training model according to size to obtain a scaling factor sequence list; Initially, a continuous preset number of the scaling factors is determined as the preset interval.

所述模型训练模块200还用于将所述N张训练图像输入至所述第二训练模型进行训练，得到中间检测模型；计算所述中间检测模型的平均精度均值；若所述平均精度均值不小于预设阈值，则确定所述中间检测模型为所述目标检测模型。Themodel training module 200 is further configured to input the N training images into the second training model for training to obtain an intermediate detection model; calculate the average precision of the intermediate detection model; If it is less than the preset threshold, the intermediate detection model is determined to be the target detection model.

所述模型训练模块200还用于若所述平均精度均值小于预设阈值，则将所述中间检测模型确定为所述第一训练模型，并返回执行所述对所述第一训练模型进行稀疏化和剪枝训练，得到第二训练模型。Themodel training module 200 is further configured to determine the intermediate detection model as the first training model if the average precision is less than a preset threshold, and return to executing the sparseness of the first training model. After pruning and pruning training, a second training model is obtained.

请参见图11，本申请一个实施例提供了一种目标识别装置20，所述装置包括：图像获取模块300、目标识别模块400和目标确定模块500。Referring to FIG. 11 , an embodiment of the present application provides atarget recognition apparatus 20 . The apparatus includes: animage acquisition module 300 , a target recognition module 400 , and a target determination module 500 .

图像获取模块300用于获取待识别图像；Theimage acquisition module 300 is used to acquire the image to be recognized;

目标识别模块400用于将所述待识别图像输入至所述目标检测模型，通过所述目标检测模型对所述待识别图像进行目标识别处理，其中，所述目标检测模型通过如上所述的目标检测模型训练方法训练得到；The target recognition module 400 is configured to input the to-be-recognized image into the target detection model, and perform target recognition processing on the to-be-recognized image through the target detection model, wherein the target detection model passes the target as described above. The detection model training method is trained;

目标确定模块500用于根据所述目标检测模型的目标识别处理结果确定目标。The target determination module 500 is configured to determine the target according to the target recognition processing result of the target detection model.

关于所述目标检测模型训练装置10和所述目标识别装置20的具体限定可以参见上文中对于目标检测模型训练和目标识别方法的限定，在此不再赘述。上述所述目标检测模型训练装置10和所述目标识别装置20中的各个模块可全部或部分通过软件、硬件及其组合来实现。上述各模块可以硬件形式内嵌于或独立于计算机设备中的处理器中，也可以以软件形式存储于计算机设备中的存储器中，以便于处理器调用执行以上各个模块对应的操作。For the specific limitations of the target detectionmodel training device 10 and thetarget recognition device 20, reference may be made to the above limitations on target detection model training and target recognition methods, which will not be repeated here. All or part of the modules in the target detectionmodel training device 10 and thetarget recognition device 20 can be implemented in whole or in part by software, hardware, and combinations thereof. The above modules can be embedded in or independent of the processor in the computer device in the form of hardware, or stored in the memory in the computer device in the form of software, so that the processor can call and execute the operations corresponding to the above modules.

在一个实施例中，提供了一种计算机设备，包括：包括存储器和处理器，所述存储器存储有计算机程序，所述处理器执行所述计算机程序时实现如下步骤：In one embodiment, a computer device is provided, comprising: a memory and a processor, the memory stores a computer program, and the processor implements the following steps when executing the computer program:

对M张样本图像进行数据增广处理，得到N张训练图像，其中M、N均为不小于1的正整数，且M＜N；Perform data augmentation processing on M sample images to obtain N training images, where M and N are both positive integers not less than 1, and M<N;

将所述N张训练图像输入至初始检测模型进行训练，得到第一训练模型；The N training images are input into the initial detection model for training to obtain the first training model;

对所述第一训练模型进行稀疏化和剪枝训练，得到第二训练模型；Perform sparse and pruning training on the first training model to obtain a second training model;

将所述N张训练图像输入至所述第二训练模型进行训练，得到目标检测模型。The N training images are input into the second training model for training to obtain a target detection model.

在一个实施例中，处理器执行计算机程序时还实现：对所述M张样本图像进行切图数据增广处理，得到N1张第一训练图像，其中N1为不小于1的正整数，且M＜N1＜N；对所述M张样本图像进行HSV色域变化数据增广处理，得到N2张第二训练图像，其中N2为不小于1的正整数，且M＜N2＜N；对所述M张样本图像进行马赛克数据增广处理，得到N3张第三训练图像，其中N1为不小于1的正整数，且M＜N3＜N。In one embodiment, when the processor executes the computer program, the processor further implements: performing image-cutting data augmentation processing on the M sample images to obtain N1 first training images, where N1 is a positive integer not less than 1, and M <N1<N; perform HSV color gamut change data augmentation processing on the M sample images to obtain N2 second training images, where N2 is a positive integer not less than 1, and M<N2<N; The M sample images are subjected to mosaic data augmentation processing to obtain N3 third training images, where N1 is a positive integer not less than 1, and M<N3<N.

在一个实施例中，处理器执行计算机程序时还实现：对对所述M张样本图像进行第一切图处理，得到N11张第一切图图像，其中N11为不小于1的正整数，且M＜N11＜N1；对所述M张样本图像进行第二切图处理，得到N12张第二切图图像，其中N12为不小于1的正整数，且M＜N12＜N1，所述第二切图处理与所述第一切图切图方式不同。In one embodiment, when the processor executes the computer program, the processor further implements: performing the first cut-through processing on the M sample images to obtain N11 first cut-through images, where N11 is a positive integer not less than 1, and M<N11<N1; perform the second cut image processing on the M sample images to obtain N12 second cut images, wherein N12 is a positive integer not less than 1, and M<N12<N1, the second cut image is obtained. The cutting process is different from the first cutting method.

在一个实施例中，处理器执行计算机程序时还实现：对按照预设模型对所述M张样本图像进行切图处理，得到M1张切图图像；对所述M1张切图图像进行筛选，去除无目标图像，得到所述N1张第一训练图像，所述无目标图像是指图像中未出现目标的图像。In one embodiment, when the processor executes the computer program, it further implements: performing image-slicing processing on the M sample images according to the preset model, to obtain M1 sliced images; screening the M1 sliced images, The non-target images are removed to obtain the N1 first training images, and the non-target images refer to the images in which the target does not appear.

在一个实施例中，处理器执行计算机程序时还实现：对确定预设区间，所述预设区间是指预设的缩放因子数值范围；去除所述第一训练模型中所述预设区间内的缩放因子对应的通道，得到所述第二训练模型。In one embodiment, when the processor executes the computer program, it further implements: determining a preset interval, where the preset interval refers to a preset scaling factor numerical range; removing the preset interval in the first training model The channel corresponding to the scaling factor is obtained to obtain the second training model.

在一个实施例中，处理器执行计算机程序时还实现：对所述第一训练模型中的多个缩放因子按照大小进行排序，得到缩放因子序列表；在所述缩放因子序列表中，从最小数值的所述缩放因子开始，将连续的预设数量的所述缩放因子确定为所述预设区间。In one embodiment, when the processor executes the computer program, the processor further implements: sorting the plurality of scaling factors in the first training model according to size to obtain a scaling factor sequence list; Starting from the scaling factor of the numerical value, a continuous preset number of the scaling factors is determined as the preset interval.

在一个实施例中，处理器执行计算机程序时还实现：将所述N张训练图像输入至所述第二训练模型进行训练，得到中间检测模型；计算所述中间检测模型的平均精度均值；若所述平均精度均值不小于预设阈值，则确定所述中间检测模型为所述目标检测模型。In one embodiment, when the processor executes the computer program, it further implements: inputting the N training images into the second training model for training to obtain an intermediate detection model; calculating the average precision of the intermediate detection model; if If the average precision is not less than a preset threshold, the intermediate detection model is determined to be the target detection model.

在一个实施例中，处理器执行计算机程序时还实现：若所述平均精度均值小于预设阈值，则将所述中间检测模型确定为所述第一训练模型，并返回执行所述对所述第一训练模型进行稀疏化和剪枝训练，得到第二训练模型。In one embodiment, when the processor executes the computer program, the processor further implements: if the average precision value is less than a preset threshold, determining the intermediate detection model as the first training model, and returning to execute the pairing of the The first training model is sparse and pruned to obtain the second training model.

在一个实施例中，处理器执行计算机程序时还实现：In one embodiment, when the processor executes the computer program, it further implements:

获取待识别图像；Get the image to be recognized;

将所述待识别图像输入至所述目标检测模型，通过所述目标检测模型对所述待识别图像进行目标识别处理，其中，所述目标检测模型通过如上所述的目标检测模型训练方法训练得到；Input the to-be-recognized image into the target detection model, and perform target recognition processing on the to-be-recognized image by the target detection model, wherein the target detection model is obtained by training the above-mentioned target detection model training method ;

根据所述目标检测模型的目标识别处理结果确定目标。The target is determined according to the target recognition processing result of the target detection model.

在一个实施例中，提供了一种计算机可读存储介质，其上存储有计算机程序，所述计算机程序被处理器执行时实现如下步骤：In one embodiment, a computer-readable storage medium is provided, on which a computer program is stored, and when the computer program is executed by a processor, the following steps are implemented:

对M张样本图像进行数据增广处理，得到N张训练图像，其中M、N均为不小于1的正整数，且M＜N；Perform data augmentation processing on M sample images to obtain N training images, where M and N are both positive integers not less than 1, and M<N;

将所述N张训练图像输入至初始检测模型进行训练，得到第一训练模型；The N training images are input into the initial detection model for training to obtain the first training model;

对所述第一训练模型进行稀疏化和剪枝训练，得到第二训练模型；Perform sparse and pruning training on the first training model to obtain a second training model;

将所述N张训练图像输入至所述第二训练模型进行训练，得到目标检测模型。The N training images are input into the second training model for training to obtain a target detection model.

在一个实施例中，计算机程序被处理器执行时还实现：对所述M张样本图像进行切图数据增广处理，得到N1张第一训练图像，其中N1为不小于1的正整数，且M＜N1＜N；对所述M张样本图像进行HSV色域变化数据增广处理，得到N2张第二训练图像，其中N2为不小于1的正整数，且M＜N2＜N；对所述M张样本图像进行马赛克数据增广处理，得到N3张第三训练图像，其中N1为不小于1的正整数，且M＜N3＜N。In one embodiment, when the computer program is executed by the processor, it further implements: performing image-cut data augmentation processing on the M sample images to obtain N1 first training images, where N1 is a positive integer not less than 1, and M<N1<N; perform HSV color gamut change data augmentation processing on the M sample images to obtain N2 second training images, where N2 is a positive integer not less than 1, and M<N2<N; The M sample images are subjected to mosaic data augmentation processing to obtain N3 third training images, where N1 is a positive integer not less than 1, and M<N3<N.

在一个实施例中，计算机程序被处理器执行时还实现：对所述M张样本图像进行第一切图处理，得到N11张第一切图图像，其中N11为不小于1的正整数，且M＜N11＜N1；对所述M张样本图像进行第二切图处理，得到N12张第二切图图像，其中N12为不小于1的正整数，且M＜N12＜N1，所述第二切图处理与所述第一切图的切图方式不同。In one embodiment, when the computer program is executed by the processor, it further implements: performing the first cut-through processing on the M sample images to obtain N11 first cut-through images, wherein N11 is a positive integer not less than 1, and M<N11<N1; perform the second cut image processing on the M sample images to obtain N12 second cut images, wherein N12 is a positive integer not less than 1, and M<N12<N1, the second cut image is obtained. The cutting process is different from the cutting method of the first cutting.

在一个实施例中，计算机程序被处理器执行时还实现：对按照预设模型对所述M张样本图像进行切图处理，得到M1张切图图像；对所述M1张切图图像进行筛选，去除无目标图像，得到所述N1张第一训练图像，所述无目标图像是指图像中未出现目标的图像。In one embodiment, when the computer program is executed by the processor, it further implements: performing image-slicing processing on the M sample images according to a preset model, to obtain M1 sliced images; screening the M1 sliced images , remove the non-target images to obtain the N1 first training images, and the non-target images refer to the images in which the target does not appear in the images.

在一个实施例中，计算机程序被处理器执行时还实现：对确定预设区间，所述预设区间是指预设的缩放因子数值范围；去除所述第一训练模型中所述预设区间内的缩放因子对应的通道，得到所述第二训练模型。In one embodiment, when the computer program is executed by the processor, it further implements: determining a preset interval, where the preset interval refers to a preset scale factor numerical range; removing the preset interval in the first training model The channel corresponding to the scaling factor within is obtained to obtain the second training model.

在一个实施例中，计算机程序被处理器执行时还实现：对所述第一训练模型中的多个缩放因子按照大小进行排序，得到缩放因子序列表；在所述缩放因子序列表中，从最小数值的所述缩放因子开始，将连续的预设数量的所述缩放因子确定为所述预设区间。In one embodiment, when the computer program is executed by the processor, the computer program further implements: sorting a plurality of scaling factors in the first training model according to size to obtain a scaling factor sequence list; in the scaling factor sequence list, from Starting from the scaling factor of the smallest value, a continuous preset number of the scaling factors is determined as the preset interval.

在一个实施例中，计算机程序被处理器执行时还实现：将所述N张训练图像输入至所述第二训练模型进行训练，得到中间检测模型；计算所述中间检测模型的平均精度均值；若所述平均精度均值不小于预设阈值，则确定所述中间检测模型为所述目标检测模型。In one embodiment, when the computer program is executed by the processor, it further implements: inputting the N training images into the second training model for training to obtain an intermediate detection model; calculating the average precision of the intermediate detection model; If the average precision is not less than a preset threshold, the intermediate detection model is determined to be the target detection model.

在一个实施例中，计算机程序被处理器执行时还实现：若所述平均精度均值小于预设阈值，则将所述中间检测模型确定为所述第一训练模型，并返回执行所述对所述第一训练模型进行稀疏化和剪枝训练，得到第二训练模型。In one embodiment, when the computer program is executed by the processor, it further implements: if the average precision value is less than a preset threshold, determining the intermediate detection model as the first training model, and returning to executing the pairing of all Perform sparsification and pruning training on the first training model to obtain a second training model.

在一个实施例中，计算机程序被处理器执行时还实现：In one embodiment, the computer program, when executed by the processor, further implements:

获取待识别图像；Get the image to be recognized;

将所述待识别图像输入至所述目标检测模型，通过所述目标检测模型对所述待识别图像进行目标识别处理，其中，所述目标检测模型通过如上所述的目标检测模型训练方法训练得到；Input the to-be-recognized image into the target detection model, and perform target recognition processing on the to-be-recognized image by the target detection model, wherein the target detection model is obtained by training the above-mentioned target detection model training method ;

根据所述目标检测模型的目标识别处理结果确定目标。The target is determined according to the target recognition processing result of the target detection model.

以上实施例的各技术特征可以进行任意的组合，为使描述简洁，未对上述实施例中的各个技术特征所有可能的组合都进行描述，然而，只要这些技术特征的组合不存在矛盾，都应当认为是本说明书记载的范围。The technical features of the above embodiments can be combined arbitrarily. In order to make the description simple, all possible combinations of the technical features in the above embodiments are not described. However, as long as there is no contradiction in the combination of these technical features It is considered to be the range described in this specification.

以上所述实施例仅表达了本申请的几种实施方式，其描述较为具体和详细，但并不能因此而理解为对发明专利范围的限制。应当指出的是，对于本领域的普通技术人员来说，在不脱离本申请构思的前提下，还可以做出若干变形和改进，这些都属于本申请的保护范围。因此，本申请专利的保护范围应以所附权利要求为准。The above-mentioned embodiments only represent several embodiments of the present application, and the descriptions thereof are specific and detailed, but should not be construed as a limitation on the scope of the invention patent. It should be pointed out that for those skilled in the art, without departing from the concept of the present application, several modifications and improvements can be made, which all belong to the protection scope of the present application. Therefore, the scope of protection of the patent of the present application shall be subject to the appended claims.

Claims (13)

1. A method for training an object detection model, the method comprising:

carrying out data amplification processing on the M sample images to obtain N training images, wherein M, N are positive integers not less than 1, and M is less than N;

inputting the N training images into an initial detection model for training to obtain a first training model;

performing sparsification and pruning training on the first training model to obtain a second training model;

and inputting the N training images into the second training model for training to obtain a target detection model.

2. The method of training an object detection model according to claim 1, wherein the N training images include: a first training image, a second training image, and a third training image; the data amplification processing is carried out on the M sample images to obtain N training images, and the method comprises the following steps:

carrying out map cutting data augmentation processing on the M sample images to obtain N1 first training images, wherein N1 is a positive integer not less than 1, and M is more than N1 and more than N;

performing HSV color gamut change data augmentation processing on the M sample images to obtain N2 second training images, wherein N2 is a positive integer not less than 1, and M is more than N2 and more than N;

and performing mosaic data amplification processing on the M sample images to obtain N3 third training images, wherein N1 is a positive integer not less than 1, and M < N3 < N.

3. The method of claim 2, wherein the N1 first training images include a first cutout image and a second cutout image; the image cutting data augmentation processing is carried out on the M sample images to obtain N1 first training images, and the image cutting data augmentation processing method comprises the following steps:

performing first cut processing on the M sample images to obtain N11 first cut images, wherein N11 is a positive integer not less than 1, and M < N11 < N1;

and performing second map cutting processing on the M sample images to obtain N12 second map cutting images, wherein N12 is a positive integer not less than 1, and M < N12 < N1, and the second map cutting processing is different from the first map cutting processing.

4. The method for training the target detection model according to claim 2, wherein the step of performing the cut-map data augmentation processing on the M sample images to obtain N1 first training images comprises:

carrying out image cutting processing on the M sample images according to a preset model to obtain M1 image cutting images;

and screening the M1 cut image, and removing the non-target image to obtain the N1 first training images, wherein the non-target image is an image without a target in the image.

5. The method for training the target detection model according to claim 1, wherein the performing the thinning and pruning training on the first training model to obtain a second training model comprises:

determining a preset interval, wherein the preset interval refers to a preset zoom factor numerical range;

and removing the channel corresponding to the scaling factor in the preset interval in the first training model to obtain the second training model.

6. The method for training the target detection model according to claim 5, wherein the determining the preset interval comprises:

sorting the multiple scaling factors in the first training model according to sizes to obtain a scaling factor sequence list;

in the scaling factor sequence table, starting from the scaling factor with the minimum value, determining a continuous preset number of scaling factors as the preset interval.

7. The method for training the target detection model according to claim 1, wherein the inputting the N training images into the second training model for training to obtain the target detection model comprises:

inputting the N training images into the second training model for training to obtain an intermediate detection model;

calculating the average precision mean value of the intermediate detection model;

and if the average precision mean value is not less than a preset threshold value, determining the intermediate detection model as the target detection model.

8. The method of training an object detection model according to claim 7, further comprising:

and if the average precision mean value is smaller than a preset threshold value, determining the intermediate detection model as the first training model, and returning to execute the thinning and pruning training of the first training model to obtain a second training model.

9. An object detection model training apparatus, characterized in that the apparatus comprises:

the data amplification module is used for performing data amplification processing on the M sample images to obtain N training images, wherein M, N are positive integers not less than 1, and M is less than N;

the model training module is used for inputting the N training images into an initial detection model for training to obtain a first training model; performing sparsification and pruning training on the first training model to obtain a second training model; and inputting the N training images into the second training model for training to obtain a target detection model.

10. A method of object recognition, comprising:

acquiring an image to be identified;

inputting the image to be recognized into the target detection model, and performing target recognition processing on the image to be recognized through the target detection model, wherein the target detection model is obtained by training through the target detection model training method according to any one of claims 1 to 8;

and determining a target according to a target identification processing result of the target detection model.

11. An object recognition apparatus, characterized in that the apparatus comprises:

the image acquisition module is used for acquiring an image to be identified;

a target recognition module, configured to input the image to be recognized into the target detection model, and perform target recognition processing on the image to be recognized through the target detection model, where the target detection model is obtained by training through the target detection model training method according to any one of claims 1 to 8;

and the target determining module is used for determining a target according to a target identification processing result of the target detection model.

12. A computer device comprising a memory and a processor, the memory storing a computer program, characterized in that the processor realizes the steps of the method of any one of claims 1 to 9 when executing the computer program.

13. A computer-readable storage medium, on which a computer program is stored, which, when being executed by a processor, carries out the steps of the method of any one of claims 1 to 9.

Images