Optimization method of image recognition algorithm, image recognition system

Abstract

The present invention discloses an optimization method and an image recognition system for an image recognition algorithm. The optimization method for the image recognition algorithm includes: step 1, training an initial algorithm model based on training data less than a first preset number; step 2, using the initial algorithm model as the algorithm model in use; step 3, inputting the image to be recognized into the algorithm model in use, and the algorithm model in use outputs an image that does not meet the requirements; step 4, manually performing a secondary judgment on the image that does not meet the requirements output by the algorithm model in use, and if it is an inference error, re-labeling the image with the inference error, and accumulating to a second preset number, processing it as training data to train the algorithm model currently in use, and the algorithm model after training is put into use as the algorithm model in use, and returning to step 3. The present invention realizes a method for automatic iterative upgrade of algorithms based on small data sets.

Description

Technical Field

[0001] The present invention relates to the technical field of image recognition algorithms, and in particular to an optimization method for an image recognition algorithm. Background Art

[0002] Image recognition technology is currently widely used, such as face recognition, traffic violation recognition, etc. Safety production is the lifeline of enterprise development, and traditional safety monitoring methods have blind spots and delays, making it difficult to detect and respond to potential risks in a timely manner. AI image recognition algorithms can effectively improve risk monitoring efficiency and ensure enterprise safety production by utilizing advanced computer vision and deep learning technologies.

[0003] AI image recognition algorithms are usually deployed on local servers or edge servers. They perform intelligent identification and early warning by analyzing image data collected by front-end surveillance cameras, drones and other equipment, and detect unsafe behaviors and conditions in the first place, thus achieving advance warning.

[0004] Currently, the application scenarios of image recognition algorithms can include the following application scenarios.

[0005] Construction safety: At construction sites, AI image recognition algorithms can be used to monitor dangerous work behaviors, and timely detect and warn of irregular behaviors such as failure to wear safety helmets and reflective clothing as required, as well as abnormal situations such as fireworks and intrusions by strangers, thereby improving safe operation management on construction sites.

[0006] Industrial safety production: In the production process of chemical plants, food processing plants, power plants and other factories, AI image recognition algorithms are used to monitor the operating status of equipment, employee operating specifications, factory / workshop safety and other dynamics, to promptly detect abnormal behaviors and situations and avoid the occurrence of safety accidents.

[0007] Mine safety: In coal and non-coal mining operations, the use of AI image recognition algorithms can help build an intelligent video management system to more efficiently monitor the safety status of people, vehicles, and objects on site, such as cross-border mining, unauthorized leaving the post, etc., and improve the level of safety management.

[0008] Road traffic safety: Use AI image recognition algorithms to monitor road traffic conditions, identify violations of non-motor vehicles and motor vehicles, as well as traffic flow, traffic accidents, etc., to improve traffic management efficiency and reduce the occurrence of traffic accidents.

[0009] Fire safety at gas stations: Strengthen safety management at gas stations through AI image recognition algorithms, and provide intelligent identification and early warning of unsafe conditions such as smoking, fireworks, and oil unloading operation specifications.

[0010] In general, AI image recognition algorithms play an important role in production safety, and can effectively improve risk monitoring efficiency and reduce the occurrence of safety accidents. However, the accuracy of current customized image recognition algorithms is not high, especially after switching to different scenes, because image recognition algorithms usually require a large amount of training data to learn patterns and features. For example, as an image recognition algorithm, if the scene of the training data is a construction safety scene, if it is switched to an industrial safety production scene, the accuracy will be reduced.

[0011] In order to solve the accuracy problem of image recognition algorithms in the prior art, it is usually necessary to collect all possible image features in different scenes, accumulate enough data, and then manually annotate them. Then, the annotated data is used as training data to train the image recognition algorithm. Therefore, many image recognition algorithms need to be based on the early big data collection process to obtain enough training data. The trained algorithm model will be deployed to the site for reasoning analysis. If the trained algorithm model does not work well, further training data will be collected to expand the amount of training data and train the algorithm model again to obtain an accurate algorithm model.

[0012] From the description of the prior art, it can be seen that the prior art requires a large amount of field data collection and manual labeling, which is very time-consuming and labor-intensive. Furthermore, for some scenarios, there are too few field samples, which makes the amount of data that can be collected too small. If it is difficult to artificially simulate samples, the accuracy of the corresponding algorithm will be very low and basically unusable.

[0013] Therefore, how to improve the efficiency of image recognition algorithms and reduce labor costs is a technical problem to be solved. Summary of the invention

[0014] In order to solve the technical problems of low efficiency and high labor cost of image recognition algorithms in the prior art, the present invention proposes an optimization method for an image recognition algorithm and an image recognition system.

[0015] The optimization method of the image recognition algorithm proposed by the present invention comprises:

[0016] Step 1, obtaining an initial algorithm model through training based on training data less than a first preset amount;

[0017] Step 2, using the initial algorithm model as the algorithm model in use;

[0018] Step 3, inputting the image to be recognized into the algorithm model being used, and the algorithm model being used outputs an image that does not meet the requirements;

[0019] Step 4: manually conduct a second judgment on the images that do not meet the requirements output by the algorithm model currently in use. If it is an inference error, the images with inference errors will be re-labeled, and after accumulating to a second preset number, they will be processed as training data to train the algorithm model currently in use. The trained algorithm model will be put into use as the algorithm model in use, and return to step 3.

[0020] Furthermore, the first preset number is greater than or equal to the second preset number.

[0021] Furthermore, the training data are images that meet the requirements, and / or the images with reasoning errors are screened in step 4, and the images with reasoning errors but actually meet the requirements are screened out, re-labeled and used as training data.

[0022] Furthermore, in step 4, the images with inference errors are uniformly annotated and used as training data.

[0023] The image recognition system proposed in the present invention adopts the optimization method of the image recognition algorithm described in the above technical solution to optimize its algorithm model; the image recognition system comprises:

[0024] A training module, which obtains the training data to obtain the initial algorithm model, and obtains the training data to obtain an iteratively updated algorithm model;

[0025] The business module puts the initial algorithm model and the iteratively updated algorithm model into use, and processes the images with reasoning errors confirmed by the audit post into training data.

[0026] Furthermore, the training module includes a training server and an algorithm training platform set up on the training server.

[0027] Furthermore, the algorithm training platform includes at least one of the following units:

[0028] A data set management unit, used for managing each batch of the training data;

[0029] A label management unit, used to manage labels of training data;

[0030] The model management unit is used to manage the initial algorithm model and each iteratively updated algorithm model.

[0031] Furthermore, the business module includes an inference device and a business platform built on the inference device.

[0032] Furthermore, the service platform includes at least one of the following units:

[0033] An inference device management unit, which manages at least one inference device on which the service platform is installed;

[0034] An inference task management unit, which manages the inference tasks of at least one image acquisition device connected to the service platform;

[0035] The reasoning result management unit manages the reasoning results.

[0036] The present invention has built an online and offline automatic iterative cycle update system to automatically update the algorithm model. By setting up an audit post, the accuracy of the algorithm model for a small amount of data can be achieved, which reduces the manpower burden and improves efficiency. The present invention constructs an intelligent inspection software system based on a small data set regression automatic iterative upgrade model to achieve a closed loop, which solves the problem that the prior art requires a large amount of manual data collection and labeling and lacks useful training data on site. The method of automatically iteratively upgrading the model based on small data set regression reduces the large amount of manpower required to collect and label data due to inaccurate algorithms. Furthermore, based on the idea of ​​reverse reasoning, the present invention automatically collects negative samples for training, solving the problem of difficult on-site data sampling. BRIEF DESCRIPTION OF THE DRAWINGS

[0037] The present invention is described in detail below with reference to the embodiments and accompanying drawings, wherein:

[0038] Figure 1 This is a flowchart of initial model training according to an embodiment of the present invention.

[0039] Figure 2 It is a schematic diagram of processing actual output results of a business system according to an embodiment of the present invention.

[0040] Figure 3 It is a closed-loop flow chart of the operation of the entire system of the present invention.

[0041] Figure 4 It is a schematic diagram of the structure of an image recognition system according to an embodiment of the present invention. DETAILED DESCRIPTION

[0042] In order to make the technical problems, technical solutions and beneficial effects to be solved by the present invention more clearly understood, the present invention is further described in detail below in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not used to limit the present invention.

[0043] Thus, a feature indicated in this specification will be used to illustrate one of the features of an embodiment of the present invention, rather than implying that each embodiment of the present invention must have the described feature. In addition, it should be noted that this specification describes many features. Although some features can be combined together to illustrate possible system designs, these features can also be used in other combinations that are not explicitly described. Thus, unless otherwise stated, the described combinations are not intended to be limiting.

[0044] In order to solve the technical problem in the prior art that training an image recognition algorithm requires massive data and manual annotation, the present invention proposes an optimization method for an image recognition algorithm. The optimization method for an image recognition algorithm mainly includes four steps.

[0045] Step 1: train an initial algorithm model based on training data less than a first preset amount. The amount of training data for the initial algorithm model can be very small, for example, 1000-3000 pieces. Compared with the millions of training data in the prior art, the amount of data is very small. This small amount of data needs to be manually labeled to form corresponding training data to train the initial algorithm model.

[0046] Step 2: Use the initial algorithm model as the algorithm model in use. Since image recognition algorithms are based on different scenarios, even an algorithm model trained with massive training data may have scenarios, patterns or features that are not considered. Therefore, the present invention will not consider the accuracy of the initial training model and put it into use first.

[0047] Step 3, input the image to be identified into the algorithm model currently in use, and the algorithm model currently in use outputs an image that does not meet the requirements. This step describes the step in which the algorithm model is officially put into use. Regardless of the initial algorithm model or the algorithm model that is subsequently iteratively updated, its input is the image to be identified in actual application, and the output is an image that does not meet the requirements. At the beginning, the algorithm model in use is the initial algorithm model. For the image obtained in the actual application scenario, after it is input into the initial algorithm model, if the image is inferred to be a correct image, that is, an image that meets the requirements, there will be no corresponding output. If the image is inferred to be an incorrect image, that is, an image that does not meet the requirements, there will be a corresponding output. Similarly, the subsequent training data from some data of actual business scenarios is used to obtain training data, and the same is true for training the iteratively updated algorithm model.

[0048] Step 4: manually conduct a second judgment on the images that do not meet the requirements output by the algorithm model currently in use. If it is an inference error, the images with inference errors will be re-labeled, and after accumulating to a second preset number, they will be processed as training data to train the algorithm model currently in use. The trained algorithm model will be put into use as the algorithm model in use, and return to step 3.

[0049] Step 4 is to make a secondary judgment on the output of the algorithm model in use. For example, an audit post is set up. The audit post manually makes a secondary judgment on the output of the algorithm model in use. Since the algorithm model in use does not output when it encounters an image that meets the requirements, and only outputs when it encounters an image that does not meet the requirements, the amount of output data is reduced. This small amount of data is manually re-judged by the audit post one by one. If it is wrong data, it can be processed to obtain further training data, and then the algorithm model currently in use is optimized and upgraded. The new algorithm model after the upgrade is put into use as the algorithm model in use. This cycle forms an automatic iteration system for optimizing the algorithm model. When the accuracy of the algorithm model currently in use is greater than or equal to the user's expected accuracy, the automatic iteration process can be stopped.

[0050] In one embodiment, the first preset number may be greater than or equal to the second preset number.

[0051] In a specific embodiment, when the above-mentioned step 4 of the present invention performs a secondary judgment on the output of the algorithm model in use, if erroneous data is encountered, the images with reasoning errors can be uniformly labeled. The unified labeling is different from the previous labeling. The data after the unified labeling is used as training data to optimize the algorithm model in use again, so as to obtain an iteratively updated algorithm model with higher accuracy.

[0052] In some scenarios, it is difficult to collect defective samples. Image recognition algorithms usually require a large amount of training data to learn patterns and features. However, many specific scenarios cannot collect enough sample data, and there is no way to simulate them on site. These are usually some scenarios that may cause production safety accidents. In a preferred embodiment, the present invention collects non-defective sample data for training. In this way, the model trained by positive samples will only detect positive results. If there is no detection result, it is a reverse result, which corresponds to some defect hidden danger data. Combined with the above self-circulating closed-loop system, the problem of not being able to collect training samples can be perfectly solved.

[0053] In the specific processing, the training data selected in step 1 can select images that meet the requirements, that is, the training data only contains correct data and does not contain wrong data. And / or, in step 4, the images with reasoning errors are screened. If there is an output result that is wrongly reasoned, but is actually an image that meets the requirements, that is, correct data, it is screened out, re-labeled, and used as training data, thereby solving the problem of not being able to collect defective data.

[0054] Taking building safety as an example, the present invention only collects images that meet the requirements, such as images of people wearing safety helmets and reflective clothing as required, and uses this data to train the initial model. Images of people not wearing safety helmets and reflective clothing as required will be output by the initial model as images that do not meet the requirements. At this time, they can be manually checked by the audit post to see if the reasoning is correct. If the reasoning is wrong, they will be re-labeled and added to the training set for further updating and training of the model.

[0055] The concept of the present invention will be further described below in conjunction with the accompanying drawings.

[0056] Figure 1 The customization flow chart of the initial algorithm model is shown. The initial algorithm model is the same as the traditional algorithm model. A small amount of data is collected on-site, such as 1000-3000 pictures, and manually annotated. The annotated data is saved in the initial data set, and then an initial algorithm model is trained, such as algorithm model 1.0.

[0057] like Figure 2 As shown, the initial algorithm model is placed in a business system (such as a smart inspection system of a business system) for reasoning, that is, it is put into actual use. An audit post will be set up in the business system to review the reasoning results. The present invention further automatically labels the review results. If the reasoning is correct, it is automatically labeled according to the reasoning results. If the reasoning is wrong, the erroneous data is classified into a common pseudo-label for labeling to generate a labeled data set in the YOLO format. Other formats are also acceptable. The present invention does not limit the specific format of the data set.

[0058] like Figure 3As shown, when the audit post collects a certain amount of data on site, such as 500 pictures, the business system automatically packages these annotated YOLO format data sets with pictures, and then sends them to the algorithm training platform through the algorithm training platform API. After receiving the training data, the algorithm training platform saves this batch of new training data to the initial data set, starts the training task, and obtains an iteratively updated algorithm model after training, such as algorithm model 2.0. The iteratively updated algorithm model is then sent to the business platform through the business platform API and put into use. The business platform distributes the iteratively updated algorithm model to the inference device using the model for automatic update. The inference device uses the updated algorithm model for inference. The inference result is further reviewed by the audit post, and the reviewed data continues to be automatically annotated and stored. When a certain amount is stored, this process is automatically repeated, thereby realizing automatic training of small data set regression and automatic iterative update of the on-site model, thereby connecting the online business platform and the offline training platform to form a closed-loop system that runs automatically.

[0059] As this closed-loop system runs longer, the accumulated data sets will increase, and the reasoning accuracy of the trained algorithm model will become higher and higher. The entire process only retains the audit post. Except for the training of the initial model, which requires manual collection of a certain amount of data, subsequent data collection and labeling do not require manual work at all.

[0060] The present invention is based on the optimization method of the above-mentioned image recognition algorithm and also protects a corresponding image recognition system. The image recognition system adopts the optimization method of the image recognition algorithm of the above-mentioned technical solution to optimize its algorithm model. The image recognition system includes a training module and a business module.

[0061] The training module obtains training data to obtain the initial algorithm model and the iteratively updated algorithm model.

[0062] The business module puts the initial algorithm model and iteratively updated algorithm model into use, and processes the images with reasoning errors confirmed by the audit post as training data.

[0063] In one embodiment, the training module includes a training server and an algorithm training platform set up on the training server.

[0064] In one embodiment, the business module includes an inference device and a business platform built on the inference device.

[0065] The algorithm training platform may include at least one of the following units.

[0066] A data set management unit, used for managing each batch of the training data;

[0067] A label management unit, used to manage labels of training data;

[0068] The model management unit is used to manage the initial algorithm model and each iteratively updated algorithm model.

[0069] The service platform includes at least one of the following units.

[0070] An inference device management unit, which manages at least one inference device on which the service platform is installed;

[0071] An inference task management unit, which manages the inference tasks of at least one image acquisition device connected to the service platform;

[0072] The reasoning result management unit manages the reasoning results.

[0073] The above-mentioned management of the present invention may be any one or more of operations such as viewing, modifying, editing, and deleting the corresponding content.

[0074] Figure 4 A specific embodiment of an image recognition system is shown.

[0075] In this embodiment, the algorithm training platform runs on the training server, and the algorithm training platform includes three management units, namely, the data set management unit, the label management unit, and the model management unit. The business platform runs on one or more inference devices, which are computers connected to various image acquisition devices, such as cameras, etc. The business platform includes three management units, namely, the inference device management unit, the inference task management unit, and the inference result management unit.

[0076] From the introduction of the image recognition system of the present invention, it can be seen that the image recognition system of the present invention is very simple to build, and can form an online and offline automatic closed-loop iterative circulation system to continuously iterate and update the algorithm model.

[0077] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions and improvements made within the spirit and principles of the present invention should be included in the protection scope of the present invention.

Claims

1. A method for optimizing an image recognition algorithm for safe production, characterized in that: include: Step 1, obtaining an initial algorithm model through training based on training data less than a first preset amount; Step 2, using the initial algorithm model as the algorithm model in use; Step 3, inputting the image to be recognized into the algorithm model being used, and the algorithm model being used outputs an image that does not meet the requirements; Step 4, manually perform a second judgment on the images that do not meet the requirements output by the algorithm model in use. If the image is an inference error, the image with the inference error is re-labeled, and after accumulating to a second preset number, it is processed as training data to train the algorithm model currently in use. The algorithm model after training is put into use as the algorithm model in use, and return to step 3; The training data in step 1 are images that meet production safety requirements.

2. The method for optimizing an image recognition algorithm according to claim 1, wherein: The first preset number is greater than or equal to the second preset number.

3. The method for optimizing an image recognition algorithm according to claim 1, wherein: In step 4, the images with inference errors are screened, and the images with inference errors but actually meeting the requirements are screened out, re-labeled and used as training data.

4. The method for optimizing an image recognition algorithm according to claim 1, wherein: In step 4, the images with inference errors are uniformly labeled and used as training data.

5. An image recognition system, characterized in that: The algorithm model of the image recognition algorithm is optimized by using the optimization method of the image recognition algorithm described in any one of claims 1 to 4; the image recognition system comprises: A training module, which acquires the training data to obtain the initial algorithm model and the iteratively updated algorithm model; The business module puts the initial algorithm model and the iteratively updated algorithm model into use, and processes the images with reasoning errors confirmed by the audit post into training data.

6. The image recognition system according to claim 5, characterized in that: The training module includes a training server and an algorithm training platform set up on the training server.

7. The image recognition system according to claim 6, characterized in that: The algorithm training platform includes at least one of the following units: A data set management unit, used for managing each batch of the training data; A label management unit, used to manage labels of training data; The model management unit is used to manage the initial algorithm model and each iteratively updated algorithm model.

8. The image recognition system according to claim 5, characterized in that: The business module includes an inference device and a business platform built on the inference device.

9. The image recognition system according to claim 8, characterized in that: The service platform includes at least one of the following units: An inference device management unit, which manages at least one inference device on which the service platform is installed; An inference task management unit, which manages the inference tasks of at least one image acquisition device connected to the service platform; The reasoning result management unit manages the reasoning results.

Images