Credibility evaluation method for recognition result of ship infrared deep learning model

Abstract

The invention discloses a credibility evaluation method for an identification result of a ship infrared deep learning model, and relates to the technical field of deep learning evaluation, and the method comprises the steps: generating a back propagation activation graph of an original ship infrared image through employing a guiding back propagation method, unimportant information is masked through image binarization processing and morphological closed operation, after the unimportant information is masked, the masked graph and the original ship infrared image are respectively input into the deep learning model to be evaluated, and the deep learning model to be evaluated is evaluated according to output results of the graph and the original ship infrared image. Judging the influence of the unimportant information on the recognition performance of the to-be-evaluated deep learning model, and further calculating a quantitative credibility evaluation value of the to-be-evaluated deep learning model; according to the scheme, objective evaluation of the credibility of a deep learning model result is achieved, a user can better understand the decision process and credibility level of the model, in addition, the user can be helped to quickly find and recognize a specific sample, and then the reliability and controllability of the model are improved.

Description

Technical Field

[0001] The present application relates to the technical field of deep learning evaluation, and particularly to a method for evaluating the credibility of the recognition results of a ship infrared deep learning model. Background Art

[0002] In the military and civilian fields, the automatic recognition technology of ship targets is crucial for applications such as maritime security, traffic monitoring, and military operations. Especially in the infrared band, due to its insensitivity to light and weather conditions, it has become an important source for obtaining ship target recognition data. In recent years, with the breakthrough progress of artificial intelligence technology in the field of image recognition, the development of ship infrared recognition systems based on deep learning has been rapidly developed.

[0003] However, most of the existing ship infrared deep learning recognition work focuses on improving the recognition accuracy. Although good improvements have been achieved in accuracy currently, the "black box" characteristics of deep learning models lead to the lack of transparency and interpretability in their decision-making processes, making it difficult to provide clear explanations and credibility evaluations. Especially in the field of military operations, the evaluation of the credibility of recognition results becomes particularly important, and there is still a lack of effective objective quantitative evaluation methods for evaluating the credibility of recognition results. In particular, in complex sea scenes and combat environments, the recognition results may be affected by various factors, such as time changes, clouds and fog, and enemy interference. In addition, when deep learning models face small samples, class imbalance, or adversarial samples, their recognition robustness may decrease significantly, which further highlights the need for evaluating the credibility of recognition results. Summary of the Invention

[0004] The purpose of the present application is to provide a method for evaluating the credibility of the recognition results of a ship infrared deep learning model, which can objectively evaluate the credibility of the results of the deep learning model, enabling users to better understand the decision-making process and credibility level of the deep learning model.

[0005] To achieve the above purpose, the present application provides the following solutions:

[0006] The present application provides a method for evaluating the credibility of the recognition results of a ship infrared deep learning model, including the following steps:

[0007] Obtain the deep learning model to be evaluated and the evaluation data set; the evaluation data set includes several original ship infrared images.

[0008] For any one of the original ship infrared images, use the guided backpropagation method to generate a backpropagation activation map; the value of any pixel point in the backpropagation activation map symbolizes the recognition importance of the pixel point in the original ship infrared image.

[0009] For any backpropagation activation map, after performing image binarization and morphological closing operations on the backpropagation activation map, an activation graphics image is obtained.

[0010] Overlay the original ship infrared image and its corresponding activation graphics image to obtain an activation masking image.

[0011] Input the original infrared image and the activation masking image into the deep learning model to be evaluated respectively, and obtain the first probability score vector corresponding to the original infrared image and the second probability score vector corresponding to the activation masking image.

[0012] According to the first probability score vector and the second probability score vector, calculate the quantization credibility evaluation value of the deep learning model to be evaluated; the higher the quantization credibility evaluation value, the better the credibility of the recognition result of the deep learning model to be evaluated.

[0013] Optionally, generate the backpropagation activation map according to the following formula:

[0014]

[0015] Among them, GB(I) is the backpropagation activation map, I is the original ship infrared image, O is the probability score vector of the deep learning model to be evaluated for the original ship infrared image, 1() is the indicator function, and when the condition in the parentheses is satisfied, its value is 1, otherwise it is 0.

[0016] Optionally, after performing image binarization and morphological closing operations on the backpropagation activation map, an activation graphics image is obtained, which specifically includes the following steps:

[0017] Perform image binarization on the backpropagation activation map to obtain a binarized image.

[0018] Perform morphological closing on the binarized image to obtain an activation graphics image.

[0019] Optionally, perform image binarization on the backpropagation activation map according to the following formula:

[0020] b i,j =θ(a i,j -t)

[0021] Among them, b i,j is the value of the pixel at coordinates (i, j) in the binarized image, θ() is the Heaviside function, a i,j is the value of the pixel at coordinates (i, j) in the backpropagation activation map, and t is the threshold selected according to the pixel value distribution.

[0022] Perform morphological closing on the binarized image according to the following formula:

[0023] ci,j = g(f(b i,j ))

[0024] where c i,j is the value of the pixel at coordinates (i, j) in the activation graphics image, g() is the dilation operation, and f() is the erosion operation.

[0025] Optionally, the original ship infrared image and its corresponding activation graphics image are overlapped to obtain an activation masking image, which specifically includes the following steps:

[0026] For any original ship infrared image, non-important pixels in the original ship infrared image are set to zero based on the corresponding activation graphics image of the original ship infrared image to obtain an intermediate infrared image.

[0027] Calculate the average pixel value of the original ship infrared image, and replace the pixels with a value of 0 in the intermediate infrared image with the average pixel value to obtain the activation masking image.

[0028] Optionally, non-important pixels in the original ship infrared image are set to zero according to the following formula:

[0029]

[0030] where d i,j is the value of the pixel at coordinates (i, j) in the intermediate infrared image, I i,j is the value of the pixel at coordinates (i, j) in the original ship infrared image, and c i,j is the value of the pixel at coordinates (i, j) in the activation graphics image.

[0031] Replace the pixels with a value of 0 in the intermediate infrared image with the average pixel value according to the following formula:

[0032]

[0033] where e i,j is the value of the pixel at coordinates (i, j) in the activation masking image, and h is the average pixel value of the original ship infrared image.

[0034] Optionally, calculate the quantization credibility evaluation value of the deep learning model to be evaluated according to the following formula:

[0035]

[0036] where Credibility is the quantization credibility evaluation value of the deep learning model to be evaluated, N is the number of original ship infrared images in the evaluation dataset, I n is the nth original ship infrared image, and E nis the activation masking image corresponding to the nth original ship infrared image, and μ is an indicator function. When the deep learning model to be evaluated predicts the original ship infrared image correctly, μ is 1; otherwise, it is 0. Model(I n ) is the first probability score vector recognized by the deep learning model to be evaluated according to the original infrared image I n , Model(I)=[score 1 ,..., score L . Model(E n ) is the second probability score vector recognized by the deep learning model to be evaluated according to the activation masking image E n , Model(E)=[mask_score 1 ,..., mask_score L . f ed (Model(I n ), Model(E n ) is the similarity distance between the first probability score vector and the second probability score vector.

[0037] Calculate the similarity distance between the first probability score vector and the second probability score vector according to the following formula:

[0038]

[0039] where L is the number of categories, score l is the probability score of category l in the first probability score vector, and mask_score l is the probability score of category l in the second probability score vector.

[0040] According to the specific embodiments provided by this application, the following technical effects are disclosed in this application:

[0041] The present application provides a method for evaluating the credibility of the recognition results of a ship infrared deep learning model. By using the guided backpropagation method to generate the backpropagation activation map of the original ship infrared image, and through image binarization processing and morphological closing operation, the unimportant information in the original ship infrared image is masked. After masking the unimportant information, the masked image and the original ship infrared image are respectively input into the deep learning model to be evaluated. According to the output results of the two, the influence of the unimportant information on the recognition performance of the deep learning model to be evaluated is judged, and then the quantitative credibility evaluation value of the deep learning model to be evaluated is calculated. The solution of the present application realizes an objective evaluation of the credibility of the deep learning model results by combining the prediction results of the model and the consistency quantification of the interpretable representation results, enabling users to better understand the decision-making process and credibility level of the model. In addition, this method can also help users quickly identify and find out the specific recognition samples to assist in analyzing the reasons for the model to make wrong decisions in specific scenarios, thereby improving the reliability and controllability of the model. BRIEF DESCRIPTION OF THE DRAWINGS

[0042] In order to more clearly illustrate the technical solutions in the embodiments of the present application or the prior art, the following will briefly introduce the drawings required for use in the embodiments. Obviously, the drawings in the following description are only some embodiments of the present application. For those of ordinary skill in the art, without creative efforts, other drawings can be obtained based on these drawings.

[0043] Figure 1 It is a flowchart of a method for evaluating the credibility of the recognition results of a ship infrared deep learning model provided by an embodiment of the present application;

[0044] Figure 2 It is a schematic flowchart of a method for evaluating the credibility of a batch of ship infrared deep learning models provided by another embodiment of the present application. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0045] The following will clearly and completely describe the technical solutions in the embodiments of the present application in conjunction with the drawings in the embodiments of the present application. Obviously, the described embodiments are only some embodiments of the present application, rather than all embodiments. Based on the embodiments in the present application, all other embodiments obtained by those of ordinary skill in the art without creative efforts belong to the scope of protection of the present application.

[0046] To make the above objects, features, and advantages of the present application more obvious and understandable, the present application will be further described in detail below in conjunction with the drawings and specific embodiments.

[0047] In an exemplary embodiment, as Figure 1As shown, a credibility evaluation method for the recognition results of a ship infrared deep learning model is provided. This method is executed by a computer device, which can be specifically executed by a computer device such as a terminal or a server alone, or jointly executed by a terminal and a server. In the embodiments of the present application, this method includes the following steps S101 to step S106:

[0048] S101. Obtain the deep learning model to be evaluated and the evaluation data set; the evaluation data set includes several original ship infrared images.

[0049] S102. For any one of the original ship infrared images, use the guided backpropagation method to generate a backpropagation activation map; the value of any pixel point in the backpropagation activation map symbolizes the recognition importance of the pixel point in the original ship infrared image. Specifically, the backpropagation activation map is generated according to the following formula:

[0050]

[0051] where \(GB(I)\) is the backpropagation activation map, \(I\) is the original ship infrared image, \(O\) is the probability score vector of the deep learning model to be evaluated for the original ship infrared image, and \(1()\) is the indicator function, whose value is 1 when the condition in the parentheses is satisfied, otherwise 0.

[0052] S103. For any one of the backpropagation activation maps, after performing image binarization processing and morphological closing operation on the backpropagation activation map, an activation graphics image is obtained. Step S103 specifically includes the following steps:

[0053] Perform image binarization processing on the backpropagation activation map to obtain a binarized image. The image binarization processing of the backpropagation activation map is performed according to the following formula:

[0054] b i,j =\(\theta(a\) i,j -t)

[0055] where \(b\) i,j is the value of the pixel at coordinates \((i, j)\) in the binarized image, \(\theta()\) is the Heaviside function, \(a\) i,j is the value of the pixel at coordinates \((i, j)\) in the backpropagation activation map, and \(t\) is the threshold selected according to the pixel value distribution.

[0056] Perform a morphological closing operation on the binarized image to obtain an activation graphics image. The morphological closing operation on the binarized image is performed according to the following formula:

[0057] c i,j =g(f(b i,j ))

[0058] where \(c\) i,jTo activate the value of the pixel at coordinates (i, j) in the graphics image, g() is the dilation operation and f() is the erosion operation.

[0059] S104. Overlap the original ship infrared image and its corresponding activated graphics image to obtain an activated masking image. Step S104 specifically includes the following steps:

[0060] For any original ship infrared image, based on the activated graphics image corresponding to the original ship infrared image, set the non-important pixels in the original ship infrared image to zero to obtain an intermediate infrared image. Set the non-important pixels in the original ship infrared image to zero according to the following formula:

[0061]

[0062] where d i,j is the value of the pixel at coordinates (i, j) in the intermediate infrared image, I i,j is the value of the pixel at coordinates (i, j) in the original ship infrared image, c i,j is the value of the pixel at coordinates (i, j) in the activated graphics image.

[0063] Calculate the average pixel value of the original ship infrared image, and replace the pixels with a value of 0 in the intermediate infrared image with the average pixel value to obtain an activated masking image. Replace the pixels with a value of 0 in the intermediate infrared image with the average pixel value according to the following formula:

[0064]

[0065] where e i,j is the value of the pixel at coordinates (i, j) in the activated masking image, and h is the average pixel value of the original ship infrared image.

[0066] S105. Input the original infrared image and the activated masking image into the deep learning model to be evaluated respectively, and obtain the first probability score vector corresponding to the original infrared image and the second probability score vector corresponding to the activated masking image.

[0067] S106. According to the first probability score vector and the second probability score vector, calculate the quantization credibility evaluation value of the deep learning model to be evaluated; the higher the quantization credibility evaluation value, the better the credibility of the recognition result of the deep learning model to be evaluated. Define Model(I n ) as the first probability score vector recognized by the deep learning model to be evaluated according to the original infrared image I n , Model(I) = [score 1 ,..., score L , Model(E n) is the second probability score vector recognized by the deep learning model to be evaluated according to the activation masked image E n Model(E)=[mask_score 1 ,...,mask_score L . First, calculate the similarity distance between the first probability score vector and the second probability score vector according to the following formula:

[0068]

[0069] where f ed (Model(I n ),Model(E n )) is the similarity distance between the first probability score vector and the second probability score vector, L is the number of categories, score l is the probability score of category l in the first probability score vector, and mask_score l is the probability score of category l in the second probability score vector.

[0070] Calculate the quantitative credibility evaluation value of the deep learning model to be evaluated according to the following formula:

[0071]

[0072] where Credibility is the quantitative credibility evaluation value of the deep learning model to be evaluated, N is the number of original ship infrared images in the evaluation dataset, I n is the nth original ship infrared image, E n is the activation masked image corresponding to the nth original ship infrared image, and μ is an indicator function. When the deep learning model to be evaluated predicts the original ship infrared image correctly, μ is 1, otherwise it is 0.

[0073] The solution of the above embodiment of the present application realizes an objective evaluation of the credibility of the deep learning model result by combining the prediction result of the model and the consistency quantization of the interpretable representation result, enabling users to better understand the decision-making process and credibility level of the model. In addition, this method can also help users quickly identify and find out the recognition specific samples to assist in analyzing the reasons for the model to make wrong decisions in specific scenarios, thereby improving the reliability and controllability of the model.

[0074] In another exemplary embodiment, the present application provides a method for evaluating the credibility of a deep learning model for ship infrared images in batches, as shown in Figure 2 and includes the following steps:

[0075] ① Train i deep learning models to be compared using a ship infrared target dataset in complex scenarios; the deep learning models obtained through training here can all be regarded as the subsequent deep learning models to be evaluated; specifically, the training cycle of each deep learning model is 100, the batch size is 32, the learning rate is 0.01, the momentum is 0.9, and the Stochastic Gradient Descent (SGD) optimizer is used to update the parameters. The division ratio of the training set, validation set, and test set is: 7:1:2.

[0076] ② Use the Guided Backpropagation method to batch generate the backpropagation activation maps of the deep learning models to be recognized for the samples in the test set. By intercepting the forward gradient of the model content, an image of the "important" feature region for the model to recognize the samples is realized. The calculation method is as follows:

[0077]

[0078] Among them, GB(I) represents the backpropagation activation map, I represents the original input image, O represents the output class score of the deep learning model (that is, the current deep learning model to be evaluated) for I, * represents element multiplication, and 1(·) is the indicator function, takes the value of 1 when, and 0 otherwise. The backpropagation activation map has the same image size as the original image, and the pixel points of the two correspond one by one. The brighter the former at a certain pixel point, the higher the recognition importance of the latter at that pixel point.

[0079] ③ Based on the pixel value distribution characteristics of the backpropagation activation image, perform image binarization operation on the backpropagation activation image to clearly divide the position of the important recognition area of the model. The calculation method is as follows:

[0080] b i,j =θ(a i,j -t)

[0081] Among them, b i,j is the pixel value after binarization, a i,j is the pixel value of the original activation image, t is the threshold selected according to the pixel value distribution, and θ is the Heaviside function.

[0082] ④ Perform an image morphological closing operation on the binarized image to further optimize the distribution of the position of the important recognition area of the model, denoted as the activation graphics image. The calculation method is as follows:

[0083] c i,j =g(f(b i,j ))

[0084] Among them, c i,j is the pixel value after the closing operation, bi,j The pixel value for the closing operation process, f(·) represents the erosion operation, and g(·) represents the dilation operation.

[0085] ⑤ "Overlap" the infrared images of ships in the test set with the corresponding activated graphics images, that is, mask the "non-important" areas identified in the samples. The corresponding calculation process is as follows:

[0086] Assume I is the original infrared image of the ship, and I i,j is the pixel value size of I at the position (i, j); C is the corresponding activated graphics image of I, and c i,j is the pixel value size of C at the position (i, j) to represent the importance degree of this point in the recognition, where c i,j ∈{0, 255}. The pixel value of the masked image is obtained by the following formula:

[0087]

[0088] Calculate the average value of the pixel values of the infrared image of the ship, so that the pixel values of the "non-important" areas identified in the masked image are equal to the average value h of the image pixel values, denoted as the activated mask image E. The pixel value e of E i,j is determined by the following formula:

[0089]

[0090] ⑥ Obtain the dual output results of the deep learning model for the original image and the activated mask image, and the Euclidean distance similarity of the dual results. The starting point of this part is that an excellent deep learning model can make correct recognition decisions based on the key important areas of the image. Therefore, the higher the robustness of the model to the irrelevant areas (masked areas), the better, that is, the lower the difference between the dual outputs of the model for the original image and the masked image, the better the credibility of the recognition feature learning. Therefore, the Euclidean distance between I and E is calculated to measure the credibility of the model for a single image.

[0091] The dual output results are expressed as follows:

[0092]

[0093] where Model i (I), Model i (E) respectively represent the probability score outputs of the i-th deep learning model for the original image I and the activated mask image E, and L represents the number of ship categories. Subsequently, the following Euclidean distance formula is used to measure the similarity between Model i (I), Model i (E):

[0094]

[0095] Use the designed reward and punishment function quantization model to quantify the credibility of the recognition results of the sample set, and obtain a numerical evaluation result. The calculation method is as follows:

[0096]

[0097] where Credibility i represents the quantization performance of the recognition credibility of the i-th deep learning model on the sample set, N is the number of images in the sample set, and μ i is an indicator function. When the deep learning model predicts correctly, μ i takes the value of 1, otherwise it is 0. It should be noted that for the samples predicted correctly, μ i gives a positive "reward", and the higher the |1 - f ed |, the better. For the samples predicted incorrectly, μ i gives a negative "punishment", and the higher the |1 - f ed |, the greater the "punishment". Therefore, the larger the Credibility i value, the better the credibility.

[0098] Regarding the credibility evaluation method for the deep learning model for ship infrared images provided above, considering that most of the existing public ship infrared datasets have limitations such as small sample size, unbalanced categories, and small scene changes, this authorization constructs 6,720 ship infrared high-fidelity simulation data with a resolution of 800 * 600 based on a domestic sea scene simulation modeling platform for technical verification. The basic information of this dataset is shown in Table 1.

[0099] Table 1 Basic Information of Ship Infrared Simulation Dataset

[0100] Information type Content and division Category Freighter, cruise ship, warship Season Spring, winter Light intensity Strong, weak Motion state Moving, stationary Air temperature 0℃、5℃、10℃、15℃、20℃ Water temperature 10℃、20℃ Distance 2 - 8 km, with a step size of 1 km Azimuth angle 0 - 360°, with a step size of 45° Elevation angle 0 - 60°, with a step size of 15°

[0101] The dataset constructed based on the above ship infrared simulation dataset is divided into a training set, a validation set, and a test set according to the ratio of 7:1:2. Four deep learning models, namely AlexNet, VggNet16, ResNet18, and MobileNet_v3_large, are used for technical verification. Each deep learning model is trained for 100 epoches, with a batch size of 32, a learning rate of 0.01, a momentum of 0.9, and the Stochastic Gradient Descent optimizer is used to update the parameters.

[0102] The comparison of the performance of the credibility index and the accuracy index performance (Accuracy, Recall, Precision, F1-score) of four types of deep learning models is given in Table 2 below. Since the number of samples in each category on the test set is equal, the Accuracy and Recall of each model are equal. Table 3 shows the credibility results of the deep learning model in different scenarios.

[0103] Table 2 Comparison of the performance of different deep learning models

[0104] Credibility Accuracy Recall Precision F1-score AlexNet 0.4211 76.93 76.93 82.05 76.91 VggNet16 0.5655 89.43 89.43 90.13 89.46 ResNet18 0.7209 96.80 96.80 96.81 96.80 MobileNet_v3_large 0.6032 91.37 91.37 91.36 91.34

[0105] Table 3 Comparison of the credibility results of different deep learning models in different scenarios

[0106]

[0107]

[0108] This embodiment proposes a credibility evaluation method for the recognition results of a shipborne infrared deep learning model. By combining the prediction results of the model and the quantification of the consistency of the interpretable representation results, an objective evaluation of the credibility of the recognition model results is realized, enabling users to better understand the decision-making process and credibility level of the model. In addition, this method can also help users quickly identify specific samples for recognition to assist in analyzing the reasons for the model to make wrong decisions in specific scenarios, thereby improving the reliability and controllability of the model.

[0109] The technical features of the above embodiments can be combined arbitrarily. For the sake of brevity of description, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, it should be considered as the scope described in this specification.

[0110] Specific examples are used in this article to elaborate on the principles and implementation methods of this application. The description of the above embodiments is only used to help understand the method and its core idea of this application; at the same time, for those of ordinary skill in the art, according to the idea of this application, there will be changes in the specific implementation methods and application scopes. In summary, the content of this specification should not be construed as a limitation to this application.

Claims

1. A credibility evaluation method for ship infrared deep learning model recognition results, characterized in that: The credibility evaluation method for the recognition result of the ship infrared deep learning model includes: Obtaining a deep learning model to be evaluated and an evaluation data set; the evaluation data set includes a plurality of original ship infrared images; For any of the original ship infrared images, a guided back propagation method is used to generate a back propagation activation map; the value of any pixel point in the back propagation activation map represents the recognition importance of the pixel point in the original ship infrared image; For any of the back-propagation activation maps, performing image binarization and morphological closing operations on the back-propagation activation map to obtain an activation graphics image; Overlapping the original ship infrared image and the corresponding activated graphics image to obtain an activated mask image; Inputting the original infrared image and the activated mask image into the deep learning model to be evaluated respectively, to obtain a first probability score vector corresponding to the original infrared image and a second probability score vector corresponding to the activated mask image; According to the first probability score vector and the second probability score vector, a quantitative credibility evaluation value of the deep learning model to be evaluated is calculated; the higher the quantitative credibility evaluation value is, the better the credibility of the recognition result of the deep learning model to be evaluated is.

2. The credibility evaluation method for ship infrared deep learning model recognition results according to claim 1 is characterized in that: The back propagation activation map is generated according to the following formula: Among them, GB(I) is the back propagation activation map, I is the original ship infrared image, O is the probability score vector of the deep learning model to be evaluated for the original ship infrared image, and 1() is the indicator function, which is 1 when the conditions in the brackets are met, otherwise it is 0.

3. The credibility evaluation method for ship infrared deep learning model recognition results according to claim 1 is characterized in that: After performing image binarization and morphological closing operations on the back propagation activation map, an activation graphics image is obtained, which specifically includes: Performing image binarization processing on the back propagation activation map to obtain a binary image; A morphological closing operation is performed on the binary image to obtain an activated graphics image.

4. The credibility evaluation method for ship infrared deep learning model recognition results according to claim 3 is characterized in that: The back propagation activation map is binarized according to the following formula: b i,j =θ(a i,j -t) Among them, b i,j is the value of the pixel with coordinates (i, j) in the binary image, θ() is the Heaviside function, and a i,j is the value of the pixel with coordinate (i, j) in the back-propagation activation map, and t is the threshold selected according to the pixel value distribution; Perform morphological closing operation on the binary image according to the following formula: c i,j =g(f(b i,j )) Among them, c i,j To activate the value of the pixel with coordinates (i, j) in the graphics image, g() is the dilation operation and f() is the erosion operation.

5. The credibility evaluation method for ship infrared deep learning model recognition results according to claim 1 is characterized in that: Overlapping the original ship infrared image and the corresponding activated graphics image to obtain an activated mask image specifically includes: For any original ship infrared image, non-important pixels in the original ship infrared image are set to zero based on the activated graphics image corresponding to the original ship infrared image to obtain an intermediate infrared image; The pixel value mean of the original ship infrared image is calculated, and the pixels with a value of 0 in the intermediate infrared image are replaced by the pixel value mean to obtain an activated mask image.

6. The credibility evaluation method for ship infrared deep learning model recognition results according to claim 5 is characterized in that: The non-important pixels in the original ship infrared image are set to zero according to the following formula: Among them, d i,j is the value of the pixel with coordinates (i, j) in the intermediate infrared image, I i,j is the value of the pixel with coordinate (i, j) in the original ship infrared image, c i,j is the value of the pixel with coordinates (i, j) in the activated graphics image; The pixels with a value of 0 in the intermediate infrared image are replaced by the mean pixel value according to the following formula: Among them, e i,j is the value of the pixel with coordinates (i, j) in the activated mask image, and h is the mean pixel value of the original ship infrared image.

7. The credibility evaluation method for ship infrared deep learning model recognition results according to claim 1 is characterized in that: The quantitative credibility evaluation value of the deep learning model to be evaluated is calculated according to the following formula: Among them, Credibility is the quantitative credibility evaluation value of the deep learning model to be evaluated, N is the number of original ship infrared images in the evaluation dataset, and I n is the nth original ship infrared image, E n is the activated mask image corresponding to the nth original ship infrared image, μ is the indicator function, when the deep learning model to be evaluated correctly predicts the original ship infrared image, μ is 1, otherwise it is 0; Model(I n ) is the deep learning model to be evaluated based on the original infrared image I n The first probability score vector obtained by identification, Model(I) = [score 1 ,...,score L ],Model(E n ) is the deep learning model to be evaluated according to the activation mask image E n The second probability score vector obtained by identification, Model(E) = [mask_score 1 ,...,mask_score L ], f ed (Model(I n ),Model(E n )) is the similarity distance between the first probability score vector and the second probability score vector; The similarity distance between the first probability score vector and the second probability score vector is calculated according to the following formula: Among them, L is the number of categories, score l is the probability score of category l in the first probability score vector, mask_score l is the probability score of category l in the second probability score vector.

Images