CN109978029A - A kind of invalid image pattern screening technique based on convolutional neural networks - Google Patents

Abstract

Translated fromChinese

本发明公开了一种基于建立卷积神经网络过滤样本模型进行无效图像样本筛选的方法，原始样本通过经由卷积神经网络构建的过滤样本模型，将大量模糊、空拍及残损等无效样本(冗余样本)图像筛选出，其余样本即为质量更好、具有更多代表性的图像样本，可作为图像分类的有效样本集。该方法最终可以实现：将无效样本通过算法筛选出，减少将大量无效样本筛选出所耗费的工时，降低人工成本。

The invention discloses a method for screening invalid image samples based on establishing a convolutional neural network filtering sample model. The original samples pass through the filtering sample model constructed by the convolutional neural network, and a large number of invalid samples (redundant, redundant, etc.) The remaining samples are image samples with better quality and more representativeness, which can be used as an effective sample set for image classification. The method can finally achieve: screening out invalid samples through an algorithm, reducing the man-hours spent in screening out a large number of invalid samples, and reducing labor costs.

Description

Translated fromChinese

一种基于卷积神经网络的无效图像样本筛选方法A Method for Screening Invalid Image Samples Based on Convolutional Neural Networks

技术领域technical field

本发明涉及机器学习领域，特别是涉及一种基于建立卷积神经网络过滤样本模型进行无效图像样本筛选的方法。The invention relates to the field of machine learning, in particular to a method for screening invalid image samples based on establishing a convolutional neural network filtering sample model.

背景技术Background technique

在在卷积神经网络进行图像分类时，需要大量图像样本进行分类作为样本库，用于构建模型。在实际的工业过程中，相对于正常工业过程而言，采集到的产品图像数据中往往存在一些无效的图像数据，例如流水线生产时会采集到的模糊图像、空拍图像以及残损图像等，这类图像对于模型的构建属于无效数据，所以该方法用于将模糊图像、空拍图像等从采集到的样本中筛选出，实现原始样本的清洗，减少筛选图像所耗费的工时，降低人工成本。When performing image classification in a convolutional neural network, a large number of image samples are required for classification as a sample library for building a model. In the actual industrial process, compared with the normal industrial process, there are often some invalid image data in the collected product image data, such as blurred images, aerial images and damaged images collected during assembly line production. Class images are invalid data for model construction, so this method is used to filter out blurred images, aerial images, etc. from the collected samples, realize the cleaning of the original samples, reduce the man-hours spent on screening images, and reduce labor costs.

发明内容SUMMARY OF THE INVENTION

本发明主要解决的技术问题是提供一种基于建立卷积神经网络过滤样本模型进行无效图像样本筛选的方法，能够花费更少的工时和人工成本将模糊图像、空拍图像等从采集到的样本中筛选出，实现原始样本的清洗。The main technical problem to be solved by the present invention is to provide a method for screening invalid image samples based on establishing a convolutional neural network filtering sample model, which can spend less man-hours and labor costs to remove blurred images, aerial images, etc. from the collected samples. screened out to achieve the cleaning of the original sample.

为解决上述技术问题，本发明采用的一个技术方案是：将未标注的部分样本图像进行模糊处理，筛选出该部分样本极度模糊的图像，进行一次模糊清理。In order to solve the above technical problem, a technical solution adopted in the present invention is to perform a blurring process on the unlabeled part of the sample image, screen out the extremely blurred image of the part of the sample, and perform a blur cleaning.

处理过后的样本经过人工分类，分为无效图像即模糊图像、空拍图像等和有效样本图像。The processed samples are manually classified into invalid images, namely blurred images, aerial shots, etc., and valid sample images.

经过过滤器处理的极度模糊图像也划分到无效图像一类，形成两种类型的图像划分，形成一个样本库。The extremely blurred images processed by the filter are also classified as invalid images, forming two types of image division and forming a sample library.

CNN算法采用该样本库构建一个包含两种分类情况的分类器模型，之后用该模型进行样本清洗，实现筛选出大量样本中的无效样本，对样本进行清理。The CNN algorithm uses the sample library to build a classifier model that includes two classification situations, and then uses the model to clean the samples to screen out invalid samples in a large number of samples and clean the samples.

本发明的有益效果是：本发明采用卷积神经网络进行样本清洗，实现筛选出大量样本中的无效样本，减少筛选图像所耗费的工时，降低人工成本。The beneficial effects of the present invention are as follows: the present invention adopts the convolutional neural network to clean the samples, realizes the screening of invalid samples in a large number of samples, reduces the man-hours spent on screening images, and reduces labor costs.

附图说明Description of drawings

图1是基于建立卷积神经网络过滤样本模型进行无效图像样本筛选的方法的流程示意图；1 is a schematic flowchart of a method for screening invalid image samples based on establishing a convolutional neural network filtering sample model;

图2是一种构建过滤样本模型的流程示意图；Fig. 2 is a kind of schematic flow chart of constructing filtering sample model;

具体实施方式Detailed ways

下面结合附图对本发明的较佳实施例进行详细阐述，以使本发明的优点和特征能更易于被本领域技术人员理解，从而对本发明的保护范围做出更为清楚明确的界定。The preferred embodiments of the present invention will be described in detail below with reference to the accompanying drawings, so that the advantages and features of the present invention can be more easily understood by those skilled in the art, and the protection scope of the present invention can be more clearly defined.

请参阅图1和图2，本发明实施例包括：Please refer to FIG. 1 and FIG. 2, the embodiments of the present invention include:

一种基于建立卷积神经网络过滤样本模型进行无效图像样本筛选的方法，运用卷积神经网络构建模型，将模型运用到采集样本中，花费更少的工时和人工成本将模糊图像、空拍图像等无效数据从采集到的样本中筛选出，实现原始样本的清洗。A method for screening invalid image samples based on the establishment of a convolutional neural network filtering sample model, using a convolutional neural network to build a model, and applying the model to the collected samples, it takes less man-hours and labor costs to filter blurred images and aerial images. and other invalid data are screened from the collected samples to realize the cleaning of the original samples.

实施例一：模糊、精化样本Example 1: Fuzzy and refined samples

(1)原始样本准备：在工业生产中采集生产产品的图像数据，以1000张作为基础数量。(1) Original sample preparation: collect image data of production products in industrial production, with 1000 images as the basic quantity.

(2)模糊过滤器：可以用opencv的cv2.Laplacian()方法，实现过滤模糊程度很高的图片，对样本实现一次清洗工作。(2) Blur filter: You can use opencv's cv2.Laplacian() method to filter pictures with a high degree of blur, and clean the samples once.

模糊、精化样本得到方式：将已准备的基础样本经过模糊过滤器，过滤出模糊程度较高的图片作为模糊样本、其余剩余样本则作为精化样本。How to get the blurred and refined samples: Pass the prepared basic samples through a blur filter, and filter out pictures with a higher degree of blur as blurred samples, and the rest of the samples as refined samples.

实施例二：样本集合Example 2: Sample Collection

(1)人工过滤：在没有过滤样本模型时，需要采取人工过滤的方式，将精化样本分类成无效样本类型和有效样本类型。(1) Manual filtering: When the sample model is not filtered, manual filtering is required to classify the refined samples into invalid sample types and valid sample types.

(2)无效样本类型：将模糊图像、空拍图像以及残损等样本图像作为无效样本类型。(2) Invalid sample type: Take blurred image, empty shot image and damaged sample image as invalid sample type.

(3)有效样本类型：其他图片清晰、特征明显的图像作为有效样本类型。(3) Valid sample types: other images with clear pictures and obvious features are used as valid sample types.

(4)样本集合：无效样本和有效样本两个类型组成一个样本集合。(4) Sample set: two types of invalid samples and valid samples form a sample set.

实施例三：构建过滤样本模型Example 3: Building a Filtered Sample Model

(1)算法：基于深度卷积神经元网络算法实现样本清洗。(1) Algorithm: The sample cleaning is realized based on the deep convolutional neural network algorithm.

(2)判断样本数量以及比例：若正负样本不均衡可以采用以下方法：(2) Judging the number and proportion of samples: If the positive and negative samples are not balanced, the following methods can be used:

过采样：增加样本中少数类样本的数量。复制少数样本或者在少数样本中加入随机噪声，干扰数据通过一定的规则生成一定的样本。Oversampling: Increase the number of minority class samples in the sample. Copy a few samples or add random noise to a few samples, and interfere with the data to generate certain samples through certain rules.

下采样：减少多数样本的数量。随机的去掉多数类样本，直到多数样本和少数样本相同。Downsampling: Reduce the number of majority samples. Randomly remove majority class samples until the majority and minority samples are the same.

(3)构建模型，当样本集合中正负样本数量合适且比例均衡时，开始计算模型，如图2的流程。(3) Build a model. When the number of positive and negative samples in the sample set is appropriate and the proportion is balanced, the model is calculated, as shown in the process of Figure 2.

准确率(accuracy)：Accuracy:

召回率(recall)是覆盖面的度量，度量有多个正例被分为正例：Recall is a measure of coverage, and the measure has multiple positive examples that are classified as positive examples:

当模型准确率ACC较低或者召回率recall较小时，计算的模型不符合要求。When the model accuracy rate ACC is low or the recall rate is small, the calculated model does not meet the requirements.

(4)将计算好的新模型放入图1流程中，进行处理。(4) Put the calculated new model into the process of Figure 1 for processing.

实施例四：满足需求的样本库Example 4: Sample library that meets requirements

(1)判断样本库是否满足需求：样本库的数量等是否满足需求。(1) Determine whether the sample library meets the requirements: whether the number of the sample library meets the requirements.

(2)循环处理：一般一次是无法得到一个数量充足的样本库，所以当样本库未达到需求时，开启循环，原始样本经过模糊过滤器过滤模糊图像，经过滤样本模型对样本图像进行分类，判断出新图像所属分类，实现样本清洗。(2) Loop processing: Generally, a sufficient number of sample libraries cannot be obtained at one time, so when the sample library does not meet the demand, the loop is opened, the original samples are filtered by the blur filter, and the sample images are classified by the filtered sample model. Determine the classification of the new image and realize sample cleaning.

Claims (8)

Translated fromChinese

1.一种基于建立卷积神经网络过滤样本模型进行无效图像样本筛选的方法，该方法包括以下步骤：1. a method for screening invalid image samples based on establishing a convolutional neural network filtering sample model, the method comprising the following steps:S1：收集工业生产中产品的图像数据，作为基础，形成原始样本集；S1: Collect image data of products in industrial production as a basis to form an original sample set;S2：构建模糊过滤器，实现过滤模糊程度高的图片；S2: Build a blur filter to filter pictures with a high degree of blur;S3：将模糊过滤器剩余图像即精化样本经过人工分类成无效样本类型和有效样本类型；S3: Manually classify the remaining images of the blur filter, that is, refined samples, into invalid sample types and valid sample types;S4：无效样本和有效样本两个类型组成一个样本集合；S4: Two types of invalid samples and valid samples form a sample set;S5：构建一个卷积神经网络过滤样本模型；S5: Build a convolutional neural network filtering sample model;S6：利用S2的模糊过滤器和S5的模型循环将大量原始样本图像进行分类，实现样本清洗，直至建好一个满足需求的样本库。S6: Use the fuzzy filter of S2 and the model cycle of S5 to classify a large number of original sample images, realize sample cleaning, until a sample library that meets the requirements is built.2.根据权利要求1所述的方法，其特征在于：在S3步骤之前，先在S2步骤中加入模糊过滤器，提前过滤掉模糊程度高的样本图像。2 . The method according to claim 1 , wherein, before step S3 , a blur filter is added in step S2 to filter out sample images with a high degree of blur in advance. 3 .3.根据权利要求1所述的方法，其特征在于：S3步骤中精化样本经过人工处理。3. The method according to claim 1, wherein in step S3, the refined sample is processed manually.4.根据权利要求3所述的方法，其特征在于：S3步骤分类成无效样本类型和有效样本类型。4. The method according to claim 3, wherein the step S3 is classified into invalid sample types and valid sample types.5.根据权利要求1所述的方法，其特征在于：S4步骤这个样本集合分类中，无效样本类型包括：模糊图像、空拍图像以及残损图像等。5 . The method according to claim 1 , wherein in the classification of the sample set in step S4 , the invalid sample types include: blurred images, aerial images, damaged images, and the like. 6 .6.根据权利要求1所述的方法，其特征在于：S5步骤中模型是基于卷积神经网络的方法模型。6. The method according to claim 1, wherein the model in step S5 is a method model based on a convolutional neural network.7.根据权利要求6所述的方法，其特征在于：S5步骤中当样本集合中正负样本数量合适且比例均衡时计算模型，当模型召回率和准确率不符合要求，则放弃该模型。7. The method according to claim 6, wherein in step S5, the model is calculated when the number of positive and negative samples in the sample set is appropriate and the proportion is balanced, and when the model recall rate and accuracy rate do not meet the requirements, the model is discarded.8.根据权利要求1所述的方法，其特征在于：当样本库未达到需求时，在S6步骤开启循环，对大量原始样本图像进行分类，实现样本清洗。8 . The method according to claim 1 , wherein when the sample library does not meet the requirements, a cycle is started in step S6 to classify a large number of original sample images to realize sample cleaning. 9 .