Optimization method of feature recognition machine learning algorithm under complex background noise conditions

Abstract

The present invention relates to the field of computer machine learning, and specifically to an optimization method for a feature recognition machine learning algorithm under complex background noise conditions. The method comprises the following steps: a. preparing an original image and adding a feature label to obtain a processed image; b. classifying the processed image according to different feature labels to obtain multiple sets of training image libraries, each training image library having at least one comparison chart; c. setting a sub-model for machine learning, wherein the sub-model corresponds to the feature label one by one; d. inputting the training image into the corresponding sub-model according to the feature label to train the sub-model; e. setting a total model, wherein the total model calls each sub-model trained in step d; f. inputting the training image into the total model to train the total model. Through the above technical scheme, the optimization speed of the machine learning image recognition algorithm under complex background and complex noise conditions is greatly accelerated, and the demand for training resources is reduced.

Description

Technical Field

[0001] The present invention relates to the field of computer machine learning, and in particular to an optimization method for a feature recognition machine learning algorithm under complex background noise conditions. Background Art

[0002] With the development of machine learning and the maturity of image recognition algorithms, this new technology is increasingly used in industrial production, especially in the classification and sorting of items and the position and posture adjustment of workpieces on assembly lines.

[0003] The aforementioned situations often require clear edge recognition of the workpiece and a simple background environment. For example, the conveyor belt that transports the workpiece is generally monochrome, and the entire production line is set up in a relatively stable factory. In this case, the optimization process of image recognition using machine learning algorithms and the advancement process of machine learning are relatively fast, and only tens of thousands of learning times are needed to accurately identify the type and posture of the workpiece on the conveyor belt.

[0004] However, there are also many fields where it is difficult to use machine learning algorithms for image recognition due to complex background environments, weak target features, blurred target edges, high background similarity and fusion.

[0005] For example, in the machining industry, the production of metal workpieces requires many processes. In the prior art, image recognition technology has been added to many of these processes for production control and defect screening. However, in the process steps of pickling, oxidation, liquid carburizing, etc., image recognition technology is difficult to extract key workpiece contours or defect features. Since the background noise environment is complex and changeable in a complex liquid environment, it is very difficult to train a machine learning algorithm for identifying defect features.

[0006] For example, in underwater salvage operations, the underwater environment is complex and the surface of the salvaged artifacts is covered with silt, which makes it difficult for the machine's image recognition algorithm to be trained to the level of accurately identifying the salvaged objects.

[0007] The reasons why image recognition is difficult to apply in complex environments are: on the one hand, the amount of sample data used to train machine learning algorithms and improve algorithm accuracy is limited; on the other hand, due to problems such as complex background environment, excessive noise, and difficulty in feature extraction, the advancement of machine learning algorithms is slow. Summary of the invention

[0008] The content of this application is used to introduce concepts in a brief form, which will be described in detail in the detailed implementation section below. The content of this application is not intended to identify the key features or essential features of the technical solution claimed for protection, nor is it intended to limit the scope of the technical solution claimed for protection.

[0009] In view of the problem that the machine learning mathematical algorithm for image recognition of complex background environments and weak feature targets in the prior art is difficult to advance, and the recognition accuracy is difficult to meet the actual application requirements, the present invention provides an optimization method for a feature recognition machine learning algorithm under complex background noise conditions, comprising the following steps:

[0010] a. Prepare the original image for training the algorithm, pre-process the original image and add feature labels to obtain the processed image;

[0011] b. Classify the processed images according to different feature labels to obtain multiple sets of training image libraries, where images with the same feature labels are grouped together; each training image library has at least one comparison image, and perform manual feature recognition on the comparison image;

[0012] c. Set the sub-model of machine learning. Multiple sub-models are set according to the types of feature labels, and the sub-models correspond to the feature labels one by one.

[0013] d. Input the training image into the corresponding sub-model according to the feature label and train the sub-model;

[0014] e. Set the overall model, which calls each sub-model trained in step d;

[0015] f. Input the training images into the overall model to train the overall model.

[0016] In some implementations, in step a, multiple different feature labels may be added to the same original image to obtain a training image with multiple different feature labels.

[0017] In some implementations, training images with multiple labels are classified into different image groups according to the feature labels they carry in step b.

[0018] In some embodiments, in step d, the training of each sub-model is independent of each other, and the training of multiple sub-models is performed in parallel.

[0019] In some embodiments, the training of the sub-model in step d includes:

[0020] d1. Set the initial environment of the sub-model; import any training image into the sub-model;

[0021] d2. Obtain the training set image model feature extraction result A; repeat N times to obtain set A N , the set A N Divide into training set B N and the verification set C N , where B N , C N Yes A NA subset of

[0022] Using C N To B N Verify and obtain the set D that satisfies the verification result N , D N It is B N A subset of

[0023] d3. Judgment D N Whether the quantity meets the set requirements,

[0024] If satisfied, proceed to the next step; if not satisfied, return to step d2;

[0025] d4. Determine whether the sub-model training has reached the set training times and whether all the training images in the training image library have been used;

[0026] If either answer is yes, proceed to the next step; otherwise, return to step d1 and replace the training image;

[0027] d5. Import the comparison image into the trained sub-model to simulate and verify the feature recognition effect of the sub-model;

[0028] If the verification requirements are met, the process ends and the trained sub-model is obtained; otherwise, the training times are reset and the process returns to step d1.

[0029] In some embodiments, step d5 performs simulation verification on the feature recognition effect of the sub-model.

[0030] First, the recognition result of the sub-model on the contrast image features is used to set n subdivision grayscale thresholds;

[0031] Compare the feature detection results under each subdivision grayscale threshold with its actual feature position to obtain the detection probability P d and false alarm probability P fa ; The feature recognition accuracy of the sub-model is calculated to be Q, and it is determined whether the recognition accuracy Q meets the set standard;

[0032] in:

[0033]

[0034] The variable that affects the correct detection of features by the sub-model is the background noise, which conforms to a one-dimensional normal random variable. 2 );

[0035]

[0036]

[0037] V is the grayscale threshold of any subdivision.

[0038] In some embodiments, in step f, the training of the overall model is first performed by shallow training and then by deep training;

[0039] Shallow training imports training images together with feature labels into the overall model, and the overall model can recognize feature labels and use them as a basis to accelerate machine learning training;

[0040] Deep training imports training images without feature labels into the overall model, simulates actual application scenarios, and establishes the overall model's recognition of background types in training images and the overall model's call to corresponding sub-models.

[0041] In some embodiments, the training process of the overall model in step f includes the following specific steps:

[0042] f1. Set the initial environment of the overall model and the calling relationship between the overall model and the sub-model;

[0043] f2. Perform shallow training, including importing training images with different feature labels into the overall model under labeled conditions for machine learning training;

[0044] f3. Determine whether the total number of model training times reaches the set number of training times.

[0045] If the set number of training times is reached, proceed to the next step, otherwise return to step f2;

[0046] f4. Evaluation of shallow training effect,

[0047] Import the comparison graph including the feature labels into the trained overall model, and simulate and verify the feature recognition effect of the sub-model;

[0048] If the verification requirements are met, proceed to the next step, otherwise reset the number of training times and return to step f2;

[0049] f5. Perform deep training, including importing training images into the overall model without labels for machine learning training;

[0050] f6. Determine whether the total number of model training times reaches the set number of training times.

[0051] If the set number of training times is reached, proceed to the next step, otherwise return to step f5;

[0052] f7. Evaluation of the effect of deep training,

[0053] Import the comparison graph including the feature labels into the trained overall model, and simulate and verify the feature recognition effect of the sub-model;

[0054] If the verification requirements are met, the training is completed to obtain the total model. Otherwise, the training times are reset and the process returns to step f5.

[0055] In some embodiments, the shallow training effect evaluation process in step f4 is the same as the deep training effect evaluation process in step f7.

[0056] Both of them first compare the feature recognition results of the total model to the comparison image, set n subdivision grayscale thresholds, and compare the feature detection results under each subdivision grayscale threshold with its actual feature position to obtain the detection probability P d , Correct non-discovery rate P n , false alarm probability P fa And the false alarm rate P m ; The feature recognition accuracy of the total model is calculated to be R, and whether the recognition accuracy R meets the set standard is determined;

[0057] in:

[0058]

[0059]

[0060] The variable that affects the correct detection of features by the sub-model is the background noise, which conforms to a one-dimensional normal random variable. 2 );

[0061]

[0062]

[0063] and

[0064] P d +P m =1

[0065] P fa +P n =1

[0066] V is the grayscale threshold of any subdivision, ΔV is the increase of the adjacent subdivision grayscale threshold compared to the previous one, and k is the set weight value.

[0067] In some embodiments, k∈[0,0.5].

[0068] The technical solution of the present invention significantly improves the training progress of the machine learning algorithm for complex background noise environments and reduces the sample quantity requirement.

[0069] In the present invention, the optimization of the machine learning algorithm is changed into two steps. First, the sub-model is trained, and then the total model is trained by calling the sub-model with the total model. Each sub-model is only for one feature label, that is, any sub-model is for one type of background environment. The noise characteristics of the same type of background environment have their commonalities, and the differences in noise characteristics are significantly reduced, which can promote the training of sub-models at a faster speed. Then the total model is trained, and the total model draws the final evaluation conclusion by calling the trained sub-models and integrating the evaluation results given by the sub-models.

[0070] Other features and advantages of the present invention will be described in detail in the following detailed description. BRIEF DESCRIPTION OF THE DRAWINGS

[0071] Figure 1 It is a logical block diagram of the present invention as a whole;

[0072] Figure 2 It is a logic block diagram of a sub-model part in one embodiment of the present invention;

[0073] Figure 3 It is a logic block diagram of the overall model part in one embodiment of the present invention. DETAILED DESCRIPTION

[0074] Embodiments of the present application will be described in more detail below with reference to the accompanying drawings. Although certain embodiments of the present application are shown in the accompanying drawings, it should be understood that the present application can be implemented in various forms and should not be construed as being limited to the embodiments set forth herein. On the contrary, these embodiments are provided to provide a more thorough and complete understanding of the present application. It should be understood that the drawings and embodiments of the present application are only for exemplary purposes and are not intended to limit the scope of protection of the present application.

[0075] It should also be noted that, for ease of description, only the parts related to the invention are shown in the drawings. In the absence of conflict, the embodiments and features in the embodiments of the present application can be combined with each other.

[0076] First of all, the purpose of the present invention is to quickly promote the optimization process of machine learning algorithms while saving resources as much as possible. In fact, under the condition of sufficient resources (data resources, computing resources, etc. for machine learning training), the existing technology for feature recognition of complex backgrounds can also achieve relatively good training effects, but it requires more training sample resources, and the training progress is slow, requiring strong computing power support and time support.

[0077] The best state that can be achieved by the optimization method of the machine learning image recognition algorithm of the present invention is the accuracy of human recognition of features in the image. Although there is overfitting in human recognition of features in the image, and the recognition rate of static features is not high, we do not know a better recognition method, nor do we understand how to achieve a better recognition method. Step a of the present invention adds feature labels to the original image manually, and the subsequent steps d5, f4 and f7 verify the feature recognition effect, which are all compromise results of this status quo.

[0078] In the present invention, a method for optimizing a feature recognition machine learning algorithm under complex background noise conditions is provided, and the optimization method comprises the following steps:

[0079] Step a, prepare the original image for training the algorithm, pre-process the original image and add feature labels to obtain the processed image. The original image here can be a general picture taken directly by a camera, or an infrared image taken by an infrared sensor, a radar image, or a sonar image generated underwater by a sonar device. Traditionally, the generated images are for the convenience of people to watch and extract information, but for computer programs, the above-mentioned images are data distributed in a two-dimensional plane, so there is no difference for computer programs. The specific image to be used should be comprehensively judged based on the actual working scene, the type of features to be identified, the cost, and the degree of expression of the target features.

[0080] Adding labels to the original image requires manual work. Labels are mainly determined based on the environment or background of the image. For example, two labels can be added based on the brightness of the environment. For example, if it is image recognition for underwater salvage operations, the underwater environment is a mud environment, a rock environment, a vegetation environment, etc., and these labels are added manually.

[0081] Different feature labels can also be added to the state of the water body, or different labels can be added according to the acquisition location and season of the original image.

[0082] It can be seen from the aforementioned method of manually adding feature labels that the same original image can obtain multiple feature labels at the same time. The same image can have feature labels for daytime (good lighting conditions), feature labels for rocks and underwater environments, and feature labels for the ocean bottom or artificial foundation pit bottom.

[0083] The main reason for distinguishing feature labels is that the natural environment is diverse and the background noise conditions are extremely complex. If the identified features are slightly weak and not obvious, the advancement speed of the machine learning algorithm will be extremely slow and repeatedly interfered by various background noises. Therefore, this paper uses artificial feature labels for background conditions (the feature labels are added for environmental features).

[0084] Image recognition algorithm is a rapidly developing field today. There are many mature recognition algorithms in the industry, and new and better algorithms are constantly emerging. This application mainly proposes an optimization method for accelerating the learning speed of machine learning algorithms, and does not improve the image recognition algorithm itself. The specific algorithm to be used depends on the actual usage scenario and the type of original image. For the image recognition algorithm itself, the existing mature algorithm can be used, which will not be elaborated in this article.

[0085] Prepare the original images for training in advance according to the selected recognition device type. In general, the original images need to be pre-processed. Depending on the specific machine learning algorithm used, pre-processing operations such as grayscale conversion, binarization, and cropping may be used. The specific pre-processing is given when a specific image recognition algorithm is selected. It is part of the recognition process of the selected specific image recognition algorithm itself.

[0086] As mentioned above, in step a, multiple different feature labels can be added to the same original image to obtain training images with multiple different feature labels. In step b, the training images with multiple labels will be classified into different image groups according to the feature labels they carry, and then the training images will be assigned and imported into different sub-models. The advantages of this setting are:

[0087] 1. Training images can be reused, reducing the demand for original images; thus reducing the number of manually processed images.

[0088] 2. Adapt to the actual situation in the natural environment, simulate the complexity of reality, and in the subsequent step f, multiple feature labels are required to train the overall model to recognize and reduce background noise.

[0089] Step b: classify the processed images according to different feature labels to obtain multiple groups of training image libraries, where images with the same feature labels are grouped together, and each training image library has at least one comparison image, and artificial feature recognition is performed on the comparison images.

[0090] The comparison image is a picture that has been marked with the correct position and outline of the features to be identified. Figure 1 At the beginning, the location of the feature to be identified can only be given by manual delineation. Therefore, the number of comparison images will not be large, far less than the number of training images. The comparison images are mainly used to judge the recognition accuracy of the algorithm in the subsequent process.

[0091] In the present invention, there are as many corresponding training image libraries as there are feature labels, and each image library corresponds to at least one comparison chart. The comparison chart here can be included in the corresponding training image library, or it can be one or a group of multiple comparison charts outside the training image library. Any comparison chart corresponds to one or more determined training image libraries, and the corresponding relationship is determined according to the feature labels.

[0092] Classification of processed images with feature labels into different training image libraries can be performed automatically by a computer program in a batch processing manner.

[0093] In step b, an image library is established. Any training image library corresponds to a feature label one by one, which is convenient for the subsequent steps to correctly call or import the required images. In order to evaluate the optimization effect, each training image library corresponds to several comparison pictures. In the subsequent steps, the comparison pictures are used to evaluate the training effect.

[0094] Step c: setting a sub-model for machine learning. Multiple sub-models are set according to the types of feature labels, and the sub-models correspond to the feature labels one by one.

[0095] Traditional optimization methods directly import training data into the machine learning algorithm (corresponding to the overall model of the present invention), and then optimize the training structure of the overall model itself, such as convolution, pooling, inter-layer normalization, and efficient attention weighting.

[0096] The optimization method of the present application no longer directly imports the training data into the final model, that is, the overall model, but splits the training process. First, sub-models are set one by one according to the types of feature labels, and each sub-model is trained separately for a specific feature label.

[0097] Step d: Input the training image into the corresponding sub-model according to the feature label to train the sub-model. In step a, the feature label is added to the training image by manual recognition and classification. The training images with the same feature label are all put into the same sub-model. In the sub-model training, on the one hand, the recognition of the feature is preliminarily established, and on the other hand, the recognition and noise reduction of the background noise feature of the feature label is established.

[0098] Step e: Set the overall model, and the overall model calls each sub-model trained in step d.

[0099] Step f: Input the training image into the overall model to train the overall model. The trained overall model will identify the type of background noise in the image to be identified, call different sub-models for feature recognition, and then output the recognition results of the sub-models. The specific operation mode of the overall model is continuously optimized by the machine learning algorithm.

[0100] Step d: Training of sub-model

[0101] In step d, the training of each sub-model is independent of each other, and the training of multiple sub-models is carried out in parallel. By independently training the sub-models in parallel, the training progress can be accelerated and the time can be shortened. Specifically, the following steps are included:

[0102] d1. Set the initial environment of the sub-model. Import any training image into the sub-model. First, initialize the selected algorithm. Each sub-model needs to be initialized independently.

[0103] d2. Obtain the training set image model feature extraction result A. Repeat N times to obtain set A N , the set A N Divide into training set B N and the validation set C N , where B N , C N Yes A N A subset of .

[0104] Using C N To B N Verify and obtain the set D that satisfies the verification result N , D N It is B N A subset of .

[0105] d3. Judgment D N The number of meets the set requirements. If it does, proceed to the next step; if not, return to step d2.

[0106] d4. Determine whether the sub-model training has reached the set number of training times, and determine whether all the training images in the training image library have been used. If either is true, proceed to the next step, otherwise return to step d1 and replace the training image. The training images are called sequentially here. After the training of training image X is completed, the next training image imported into the sub-model is training image X+1, until the last training image in the training image library.

[0107] d5. Import the comparison image into the trained sub-model and simulate and verify the feature recognition effect of the sub-model.

[0108] If the verification requirements are met, the trained sub-model is obtained, otherwise the training times are reset and the process returns to step d1. Setting the training times is a forced exit mechanism to avoid program dead loops. After reaching the set training times, the results will be evaluated regardless of the training level, and then the program will be terminated, the results will be output, and the training progress will be saved.

[0109] To evaluate the training effect of the sub-model, manual evaluation is actually the best method. However, the workload is too huge. Therefore, this application has made corresponding compromises and adopted the following method:

[0110] First, the sub-model sets n subdivision grayscale thresholds for the feature recognition results of the comparison image, and compares the feature detection results under each subdivision grayscale threshold with its actual feature position to obtain the detection probability P. d and false alarm probability P fa The feature recognition accuracy of the sub-model is calculated to be Q, and it is determined whether the recognition accuracy Q meets the set standard;

[0111] in:

[0112]

[0113] The variable that affects the correct detection of features by the sub-model is the background noise, which conforms to a one-dimensional normal random variable. 2 );

[0114]

[0115]

[0116] V is the grayscale threshold of any subdivision.

[0117] The difficulty in advancing the algorithm in the evaluation process of the machine learning algorithm. The core reason is that the operator cannot know in advance how the sub-model handles the background noise, and the background noise of each training image and each evaluation image is different. Therefore, the present invention adopts the method of setting n subdivision grayscale thresholds. The result of each subdivision grayscale threshold is evaluated, and then the formula is used:

[0118]

[0119] Comprehensive evaluation is performed. The number Q can be a negative number, and the larger the value, the better the recognition effect. Therefore, in step d, each evaluation can be compared with the evaluation result Q obtained from the last training. If the evaluation result Q of this time is lower than the last Q value, it can be rolled back to the state at the end of the last sub-model training.

[0120] The advantages of step d are:

[0121] 1. First, train the sub-models separately according to the types of feature labels, which can greatly reduce the difficulty of sub-model training. The background noise of the same feature label has certain common characteristics, and the background noise characteristics of training images with the same feature label are less different from each other; machine learning algorithms can easily identify the commonalities.

[0122] 2. The evaluation method of subdivided grayscale threshold is adopted to reduce the difficulty of evaluating the machine learning effect as much as possible and improve the evaluation efficiency. It is only necessary to use the formula given in the article to calculate P under each subdivided grayscale threshold. d and P fa , and then find Q.

[0123] Specific instructions for overall model training

[0124] In step f, the training of the overall model is first performed by shallow training and then by deep training.

[0125] Shallow training imports the training images together with the feature labels into the overall model, and the overall model can recognize the feature labels and use them as a basis to accelerate the training of machine learning. Deep training imports the training images without feature labels into the overall model, simulates the actual application scenario, establishes the overall model's recognition of the background types in the training images and the overall model's call to the corresponding sub-model. The specific steps include:

[0126] f1. Set the initial environment of the overall model and the calling relationship between the overall model and the sub-model;

[0127] f2. Perform shallow training, including importing training images with different feature labels into the overall model under labeled conditions for machine learning training;

[0128] f3, determine whether the total model training times reaches the set training times, if it reaches the set training times, proceed to the next step, otherwise return to step f2;

[0129] f4. Shallow training effect evaluation: import the comparison graph containing feature labels into the trained overall model to simulate and verify the feature recognition effect of the sub-model.

[0130] If the verification requirements are met, proceed to the next step, otherwise reset the number of training times and return to step f2;

[0131] f5. Perform deep training, including importing training images into the overall model without labels for machine learning training.

[0132] f6. Determine whether the total number of model training times reaches the set number of training times. If it reaches the set number of training times, proceed to the next step, otherwise return to step f5.

[0133] f7, deep training effect evaluation, import the comparison graph containing feature labels into the trained total model, and simulate and verify the feature recognition effect of the sub-model. If the verification requirements are met, the trained total model is obtained, otherwise the training times are reset and returned to step f5.

[0134] In the shallow training stage, the training images are labeled with features to assist the overall model in calling the corresponding sub-model. In this process, the overall model will be trained on how to integrate the different results fed back by different sub-models and weight the results. In the deep training stage, the overall model will be trained to recognize background features and call the appropriate sub-model.

[0135] For the shallow training effect evaluation in step f4 and the deep training effect evaluation in step f7, the processes and operations of step f4 and step f7 are exactly the same.

[0136] First, the feature recognition results of the total model are compared with the image, and n subdivision grayscale thresholds are set. The feature detection results under each subdivision grayscale threshold are compared with their actual feature positions to obtain the detection probability P. d , Correct non-discovery rate P n , false alarm probability P fa And the false alarm rate P m The feature recognition accuracy of the total model is calculated to be R, and it is determined whether the recognition accuracy R meets the set standard.

[0137] in:

[0138]

[0139]

[0140] The variable that affects the correct detection of features by the sub-model is the background noise, which conforms to a one-dimensional normal random variable. 2 );

[0141]

[0142]

[0143] and

[0144] P d +P m =1

[0145] P fa +P n =1

[0146] V is the grayscale threshold of any subdivision, ΔV is the increase of the adjacent subdivision grayscale threshold compared to the previous one, and k is the set weight value. Preferably, k∈[0, 0.5].

[0147] Through the above four formulas, we can get P at any subdivision grayscale threshold. d , P m , P fa and P n .

[0148] Advantages of the present invention:

[0149] The present invention changes the optimization of the machine learning algorithm into two steps: first, the sub-model is trained, and then the overall model is trained by calling the sub-model. Each sub-model is only for one feature label, that is, any sub-model is for one type of background environment. The noise characteristics of the same type of background environment have their commonalities, and the differences in noise characteristics are significantly reduced, which can promote the training of sub-models at a faster speed.

[0150] Then the overall model is trained. The overall model calls the trained sub-models and synthesizes the evaluation results given by the sub-models to draw the final evaluation conclusion. The training of the overall model is divided into shallow training and deep training to accelerate the correct calling relationship between the overall model and the sub-model.

[0151] The present invention provides an evaluation method for the training effect of sub-models and the total model, and also provides a specific calculation formula. Under the condition that the technical personnel cannot predict the background noise characteristics, an efficient evaluation algorithm method is proposed, which is simple to calculate, occupies less computing resources, and can run automatically without human intervention.

[0152] The above description is only some preferred embodiments of the present application and an explanation of the technical principles used. Those skilled in the art should understand that the scope of the invention involved in the embodiments of the present application is not limited to the technical solutions formed by the specific combination of the above technical features, but should also cover other technical solutions formed by any combination of the above technical features or their equivalent features without departing from the above invention concept. For example, the above features are replaced with (but not limited to) the technical features with similar functions disclosed in the embodiments of the present application to form a technical solution.

Claims

1. A method for optimizing a feature recognition machine learning algorithm under complex background noise conditions, characterized in that: The following steps are involved: a. Prepare the original image for training the algorithm, pre-process the original image and add feature labels to obtain the processed image; b. Classify the processed images according to different feature labels to obtain multiple sets of training image libraries, where images with the same feature labels are grouped together; each training image library has at least one comparison image, and perform manual feature recognition on the comparison image; c. Set the sub-model of machine learning. Multiple sub-models are set according to the types of feature labels, and the sub-models correspond to the feature labels one by one. d. Input the training image into the corresponding sub-model according to the feature label and train the sub-model; e. Set the overall model, which calls each sub-model trained in step d; f. Input the training images into the overall model to train the overall model; Step d: The training of sub-model includes: d1. Set the initial environment of the sub-model; import any training image into the sub-model; d2. Obtain the training set image model feature extraction result A; repeat N times to obtain set A N , the set A N Divide into training set B N and the verification set C N , where B N , C N Yes A N A subset of Using C N To B N Verify and obtain the set D that satisfies the verification result N , D N It is B N A subset of d3. Judgment D N Whether the number meets the set requirements, if yes, proceed to the next step, if not, return to step d2; d4. Determine whether the sub-model training has reached the set training times and whether all the training images in the training image library have been used; If either answer is yes, proceed to the next step; otherwise, return to step d1 and replace the training image; d5. Import the comparison image into the trained sub-model to simulate and verify the feature recognition effect of the sub-model; If the verification requirements are met, the process ends and the trained sub-model is obtained. Otherwise, the number of training times is reset and the process returns to step d1. In step d5, the feature recognition effect of the sub-model is simulated and verified, including: first, the recognition result of the sub-model on the feature of the comparison image is set to n subdivision grayscale thresholds; the feature detection result under each subdivision grayscale threshold is compared with its actual feature position to obtain the detection probability P d and false alarm probability P fa ; The feature recognition accuracy of the sub-model is calculated to be Q, and it is determined whether the recognition accuracy Q meets the set standard; in: The variable that affects the correct detection of features by the sub-model is the background noise, which conforms to a one-dimensional normal random variable. 2 ); V is the grayscale threshold of any subdivision.

2. The optimization method of feature recognition machine learning algorithm under complex background noise conditions according to claim 1, characterized in that: In step a, multiple different feature labels can be added to the same original image to obtain training images with multiple different feature labels.

3. The optimization method of feature recognition machine learning algorithm under complex background noise conditions according to claim 2, characterized in that: Training images with multiple labels will be classified into different image groups according to the feature labels they carry in step b.

4. The optimization method of feature recognition machine learning algorithm under complex background noise conditions according to claim 1, characterized in that: In step d, the training of each sub-model is independent of each other, and the training of multiple sub-models is carried out in parallel.

5. The optimization method of feature recognition machine learning algorithm under complex background noise conditions according to claim 1, characterized in that: Step f is to train the overall model, first performing shallow training and then deep training; Shallow training imports training images together with feature labels into the overall model, and the overall model can recognize feature labels and use them as a basis to accelerate machine learning training; Deep training imports training images without feature labels into the overall model, simulates actual application scenarios, and establishes the overall model's recognition of background types in training images and the overall model's call to corresponding sub-models.

6. The optimization method of feature recognition machine learning algorithm under complex background noise conditions according to claim 5, characterized in that: Step f The training process of the overall model includes the following specific steps: f1. Set the initial environment of the overall model and the calling relationship between the overall model and the sub-model; f2. Perform shallow training, including importing training images with different feature labels into the overall model under labeled conditions for machine learning training; f3. Determine whether the total number of model training times reaches the set number of training times. If the set number of training times is reached, proceed to the next step, otherwise return to step f2; f4. Evaluation of shallow training effect, Import the comparison graph including the feature labels into the trained overall model, and simulate and verify the feature recognition effect of the sub-model; If the verification requirements are met, proceed to the next step, otherwise reset the number of training times and return to step f2; f5. Perform deep training, including importing training images into the overall model without labels for machine learning training; f6. Determine whether the total number of model training times reaches the set number of training times. If the set number of training times is reached, proceed to the next step, otherwise return to step f5; f7. Evaluation of the effect of deep training, Import the comparison graph including the feature labels into the trained overall model, and simulate and verify the feature recognition effect of the sub-model; If the verification requirements are met, the training is completed to obtain the total model. Otherwise, the training times are reset and the process returns to step f5.

7. The optimization method of feature recognition machine learning algorithm under complex background noise conditions according to claim 6, characterized in that: The evaluation of shallow training effect in step f4 and the evaluation of deep training effect in step f7 are both: First, the feature recognition results of the total model are compared with the image, n subdivision grayscale thresholds are set, and the feature detection results under each subdivision grayscale threshold are compared with their actual feature positions to obtain the detection probability P d , Correct non-discovery rate P n , false alarm probability P fa And the false alarm rate P m ; The feature recognition accuracy of the total model is calculated to be R, and whether the recognition accuracy R meets the set standard is determined; in: The variable that affects the correct detection of features by the sub-model is the background noise, which conforms to a one-dimensional normal random variable. 2 ); and P d +P m =1 P fa +P n =1 V is the grayscale threshold of any subdivision, ΔV is the increase of the adjacent subdivision grayscale threshold compared to the previous one, and k is the set weight value.

8. The optimization method of feature recognition machine learning algorithm under complex background noise conditions according to claim 7, characterized in that: k∈[0，0.5]。

Images