A kind of the intelligent cloud plateform system and operating method of customizable algorithm of target detectionTechnical field
The present invention relates to a kind of intelligent cloud plateform system and operating method, a kind of more specific customizable algorithm of target detectionIntelligent cloud plateform system, belong to the technical field of deep learning.
Background technique
In recent years, the development of deep learning is graduallyd mature and is widely used, and is based especially on video identificationAlgorithm of target detection achieves huge progress under the development of deep learning in the past few years, and detection performance is obviously mentionedIt rises.
Target detection is a vital task of computer vision field, it can identify the multiple objects of a picture,And different objects can be oriented, provide bounding box.Target detection some occasions comparative maturity application, such as in nothingIn the application that people drives, it can detecte out the targets such as vehicle, pedestrian, traffic light；For another example in the application of safety monitoring,It can detecte face, humanoid, object etc..
The algorithm of target detection of mainstream is mainly all based on deep learning model at present, is segmented into two major classes: 1)Two-stage detection algorithm, the problem of will test are divided into two stages, first generation candidate region (regionProposals), then classify (generally also needing to position refine) to candidate region, the Typical Representative of this kind of algorithm isFasterRCNN；2) one-stage detection algorithm does not need in the region proposal stage, directly to generate the classification of objectProbability and position coordinate value, the Typical Representative of this kind of algorithm are YOLO.
There are also many disclosed data sets, such as PASCAL VOC (The PASCAL Visual Object for target detectionIt Classification) is the famous data set in one, the fields such as target detection, classification, segmentation, it includes about 10000 to haveThe picture of bounding box is for training and verifying；It is the MS COCO that Microsoft Corporation is established there are one famous data set(Common Objects in COntext) data set, for object detection task, COCO includes 80 classifications altogether, every yearThe training of contest and validation data set comprise more than 120000 pictures, the picture of the test more than 40000.
Although there is also some problems for it although algorithm of target detection has had mature application.It is most important justIt is technical threshold height, the development cycle is long.According to the characteristics of supervised learning, the exploitation of an algorithm of target detection is walked in deep learningIt suddenly include the following: mark sample, algorithm type selecting, network training, test publication, entire development process is simultaneously remarkable, needs richThe deep learning algorithm engineering teacher of rich experience can complete, and technical level is extremely difficult to for ordinary user, which limitPopularization of the algorithm of target detection in more minority's scenes.
Summary of the invention
In order to solve above-mentioned prior art problem, it is an object of that present invention to provide have to can solve ordinary user and opening the doorThe technical threshold encountered when algorithm of target detection is high, a kind of intelligence of customizable algorithm of target detection of the problem of development cycle lengthCloud platform system can be changed.
It is another object of the present invention to provide have the development procedure of algorithm of target detection can be accomplished standardization,Automation, summary can reduce technical threshold, a kind of intelligence of customizable algorithm of target detection of the technologies such as Speeding up development periodThe operating method of cloud platform system can be changed.
To achieve the goals above, the present invention is achieved by the following technical solutions:
A kind of intelligent cloud plateform system of customizable algorithm of target detection, the system include algorithm distribution platform and bandThere is the algorithm training platform of operation interface, the algorithm training platform communication link is connected to custom algorithm sample database, automatized scriptAnd sampled data end, communication link is connected to custom algorithm model, the calculation between the algorithm training platform and algorithm distribution platformMethod distribution platform communication link is connected to test client；
Wherein, the sampled data end acquisition includes the picture of target and is labeled acquisition samples pictures, the algorithmTraining platform receives the samples pictures that sampled data end uploads and formats, and it is acceptable to be converted into deep learning frameFormat, the automatized script carry out algorithm training, automatically update trained algorithm to custom algorithm model after the completion of trainingMiddle formation algorithm model, then algorithm model publication is completed by algorithm distribution platform, the algorithm distribution platform externally providesAPI service, the test client is by calling the API of algorithm distribution platform platform to verify assessment algorithm effect to realize fasterSpeed make algorithm publication or adjustment.
As an improvement the sampled data end includes deep learning image labeling tool, picture pick-up device, picture pick-up deviceAcquisition picture transfer gives deep learning tool, and deep learning tool is labeled the specific picture of needs, and the camera shooting is setStandby includes video camera, and the study image labeling tool includes yolo_mark, LabelImg.
As an improvement the acceptable format of deep learning frame includes LMDB, the deep learning frame packetInclude Caffe.
A kind of operating method of the intelligent cloud plateform system of customizable algorithm of target detection of the present invention, system operatio stepIt is rapid as follows:
S1: newly-built algorithm: user fills in the basic description information of algorithm by the operation interface of algorithm training platform；
S2: selection algorithm model: the algorithm model and net of designated pictures detection in the range of algorithm training platform is supportedNetwork structure, the algorithm model include YOLO, SSD, FasterRCNN, the network structure include VGG, ResNet,DarkNet；
S3: mark sample simultaneously uploads to platform: installation video camera is acquired the picture comprising target, passes through deep learningImage labeling tool is labeled acquisition samples pictures to the specific objective picture of needs, then samples pictures are uploaded to algorithm instructionIn the custom algorithm sample database for practicing platform；
S4: the triggering operation of completion event network training: is uploaded by samples pictures；
S5: test publication: algorithm distribution platform externally provides API service, and test client is flat by calling algorithm publicationThe API of platform carries out transmission API request, and indicate AlgID and algorithm after etc. it is to be answered, algorithm distribution platform receiveAPI request, and corresponding algorithm operation test is found according to AlgID, operation test result is finally fed back into test clientTo realize more rapidly verifying assessment algorithm effect and adjust.
As an improvement operational process includes: in the network training
A the samples pictures of mark) are downloaded from custom algorithm sample database；
B) samples pictures are formatted, change into the acceptable format of deep learning frame to meet deep learning frameThe requirement of the input data format of frame pair；
C automatized script) is called to complete algorithm training；
D) algorithm is updated into custom algorithm model after the completion of training, is used for algorithm distribution platform.
The utility model has the advantages that can solve the technical threshold height that ordinary user encounters in enabling algorithm of target detection, exploitationThe problem of period length；The development procedure of one algorithm of target detection can be accomplished to standardization, automation, summary；It can dropLow technical threshold, Speeding up development period.
Detailed description of the invention
Fig. 1 is principle of the invention structural schematic diagram.
Specific embodiment
Below in conjunction with Figure of description, the invention will be further described, but the invention is not limited to following embodiments.
A kind of intelligent cloud plateform system of customizable algorithm of target detection, the system include algorithm distribution platform and bandThere is the algorithm training platform of operation interface, the algorithm training platform communication link is connected to custom algorithm sample database, automatized scriptAnd sampled data end, communication link is connected to custom algorithm model, the calculation between the algorithm training platform and algorithm distribution platformMethod distribution platform communication link is connected to test client；
Wherein, the sampled data end acquisition includes the picture of target and is labeled acquisition samples pictures, the samplingData terminal includes deep learning image labeling tool, picture pick-up device, specifically by picture pick-up device acquisition picture transfer to deep learningTool, deep learning tool are labeled the specific picture of needs, and picture pick-up device includes video camera, learn image labeling workTool includes yolo_mark, LabelImg；
The algorithm training platform receives the samples pictures that sampled data end uploads and formats, and is converted into depthThe acceptable format of learning framework, the acceptable format of deep learning frame includes LMDB, the deep learning frame packetInclude Caffe；
The automatized script carries out algorithm training, automatically updates trained algorithm to custom algorithm mould after the completion of trainingFormation algorithm model in type, then algorithm model publication is completed by algorithm distribution platform；
The algorithm distribution platform externally provides API service, and the test client is by calling algorithm distribution platform flatThe API verifying assessment algorithm effect of platform is to realize that faster speed makes algorithm publication or adjustment.
A kind of intelligent cloud plateform system of customizable algorithm of target detection of the present invention, by presetting some mainstreamsThe algorithm model (YOLO, SSD, FastRCNN) of target detection is selected for user, at present the algorithm model of algorithm of target detection(YOLO, SSD, FastRCNN etc.) comparative maturity, and accomplished more excellent effect, user only needs the money according to oneselfSource situation selects corresponding algorithm model, and does not have to repeat entire algorithm training process.User uploads it toward algorithm training platformSamples pictures of interest, then system is known will detect for which Target Photo, can realize that supervision is learned by automatized scriptThe standardized training process of habit, therefore system is by reducing target inspection for the algorithm development process automation of entire target detectionThe exploitation threshold of method of determining and calculating.
A kind of operating method of the intelligent cloud plateform system of customizable algorithm of target detection of the present invention, system operatio stepIt is rapid as follows:
S1: newly-built algorithm: user fills in the basic description information of algorithm by the operation interface of algorithm training platform；
S2: selection algorithm model: the algorithm model and net of designated pictures detection in the range of algorithm training platform is supportedNetwork structure, the algorithm model include YOLO, SSD, FasterRCNN, the network structure include VGG, ResNet,DarkNet, the development of algorithm of target detection is relatively mature, the algorithm types based on mainstream all can accomplish compared withThe problem of excellent effect, difference is only resource overhead.User be free to put down by selection algorithm model and specified network structureThe contradiction of weighing apparatus network performance and resource overhead；
S3: mark sample simultaneously uploads to platform: installation video camera is acquired the picture comprising target, passes through deep learningImage labeling tool is labeled acquisition samples pictures to the specific objective picture of needs, then samples pictures are uploaded to algorithm instructionIn the custom algorithm sample database for practicing platform, study image labeling tool includes yolo_mark, LabelImg；
S4: the triggering operation of completion event, operational process packet in the network training network training: are uploaded by samples picturesIt includes:
A the samples pictures of mark) are downloaded from custom algorithm sample database, samples pictures as shown in Figure 1 include jpg format；
B) samples pictures are formatted, change into the acceptable format of deep learning frame to meet deep learning frameThe requirement of the input data format of frame pair；
C automatized script) is called to complete algorithm training, automatized script includes: a. production 1mdb, b: training script, c:Model publication；
D) algorithm is updated into custom algorithm model after the completion of training, is used for algorithm distribution platform；
S5: test publication: algorithm distribution platform externally provides API service, and test client is flat by calling algorithm publicationThe API of platform carries out transmission API request, and indicate AlgID and algorithm after etc. it is to be answered, algorithm distribution platform receiveAPI request, and corresponding algorithm operation test is found according to AlgID, operation test result is finally fed back into test clientTo realize more rapidly verifying assessment algorithm effect and adjust, in custom algorithm model as shown in Figure 1, AlgID 111 is correspondingModel be 111.model, the corresponding model of AlgID 222 is 222.model, and the difference AlgID corresponds to corresponding mouldType.
Finally it should be noted that present invention is not limited to the above embodiments, there can also be many variations.This field it is generalAll deformations that logical technical staff directly can export or associate from present disclosure, are considered as of the inventionProtection scope.