Interpretability method and system of deep reinforcement learning model in unmanned driving scenarios

Abstract

The present invention relates to the field of model learning, and specifically to a method and system for interpretability of a deep reinforcement learning model in an unmanned driving scenario. The method and device select a suitable simulation environment and a suitable deep reinforcement learning algorithm, obtain a converged reinforcement learning model through training, input pictures taken in an unmanned driving scenario into the reinforcement learning model, divide the features of the pictures and perform quantitative analysis of the influence of the features, calculate the influence of each feature on the model decision, and obtain the corresponding difference matrix to obtain an improved network model. This at least solves the technical problem that the prior art cannot accurately analyze the influence of each feature of the picture on the model decision.

Description

Technical Field

[0001] The present invention belongs to the field of model learning, and in particular relates to an interpretability method and system for a deep reinforcement learning model in an unmanned driving scenario. Background Art

[0002] The interpretability technology of deep reinforcement learning models in unmanned driving scenarios realizes the interpretation of opaque models in unmanned driving scenarios. It has built-in algorithms and optimization solutions to solve the problems of deep reinforcement learning models and interpretation problems in unmanned driving scenarios, and explains and visualizes the important factors in the operation and decision-making process of models in unmanned driving environments to users. Among them, deep reinforcement learning models mainly involve the selection of various deep reinforcement learning algorithms (DRL) for intelligent agents to make autonomous decisions in unmanned driving environments. Explainable artificial intelligence (Explainable AI, XAI) is an emerging field of artificial intelligence, which mainly explains and visualizes various AI algorithm models. Explainable reinforcement learning (Explainable Reinforcement Learning, XRL) is a branch of XAI technology. It interprets reinforcement learning models through a series of means, including the interpretation of deep reinforcement learning currently combined with deep learning, and visualizes them in a user-understandable text or picture format and presents them to users.

[0003] DRL is an algorithm that combines deep learning and reinforcement learning. It combines the perception ability of deep learning and the decision-making ability of reinforcement learning, and has been further developed, including value-based DRL, policy-based DRL, model-based DRL, hierarchical-based DRL, etc.

[0004] XRL is interpretable reinforcement learning, which is used to explain and visualize DRL. XRL technology is classified as follows:

[0005] (1) According to the time when information is extracted, it can be divided into: intrinsic interpretability, which is constructed to be intrinsically interpretable or interpretable in itself during training, such as decision trees; post-hoc interpretability, which is to provide an explanation for the original model by creating a second simpler model or other operations such as perturbation after training. Alternative models or saliency maps are typical.

[0006] (2) According to the scope of the explanation, it is divided into: global explanation (global interpretability) and local explanation (local interpretability). The global explanation explains the entire, general model behavior, while the local explanation provides an explanation for a specific decision. In practice, it is difficult to achieve global model interpretability, especially for models with more than a few parameters. Therefore, local interpretability can be more easily applied. Explaining the reasons for making a specific decision or a single prediction means that interpretability occurs locally. Typically, this kind of interpretability is used to generate separate explanations to explain why the model made a specific decision for an instance.

[0007] The interpretability technology of deep reinforcement learning models is a key issue in the fields of autonomous driving and computers. XRL, as a subfield of XAI, has not been widely studied. The current research direction of deep reinforcement learning is post-hoc interpretability, which includes both global interpretability and local interpretability. Although XRL started late, there are currently several typical studies in post-hoc interpretability. At the same time, the interpretability algorithms for other artificial intelligence models in XAI can also be used in the interpretation of DRL.

[0008] Post-hoc Interpretability: In 2018, Greydanus et al. proposed the Saliency Map method in an ICML paper. This is a perturbation-based method that directly perturbs the input. By performing regional Gaussian blurring and comparing the difference between the normal image and the blurred image when they pass through the network, the blurred area slides on the image, thereby traversing the entire image and obtaining multiple differences. Among them, the area with large differences plays an important role in the decision-making of the agent, and the key area of ​​DRL learning is obtained. However, existing perturbation-based methods for calculating saliency usually highlight input areas that are irrelevant to the actions taken by the agent. In 2020, the method SARFA (Specific and Relevant Feature Attribution) proposed by Nikaash et al. in an ICLR paper generates a more prominent saliency map by balancing two aspects (specificity and relevance), which capture different saliency requirements. The first records the impact of the perturbation on the relative expected return of the action to be explained. The second part weighs irrelevant features that change the relative expected return of actions other than the action to be explained. It is also possible to approximate the original black box model by training a second interpretable model. In 2016, Ribeiro et al. proposed the LIME algorithm in a SIGKDD paper, training a linear interpretable model to approximate the original classification network, thereby explaining the convolutional (CNN) classification network. At the same time, this method also uses perturbations to the image.

[0009] Existing interpretability techniques for deep reinforcement learning models are all based on perturbation techniques and training interpretable model approximation. However, the methods proposed so far have limitations and great room for improvement. In addition, the interpretation range cannot be quantified or the interpretation speed is low, which cannot provide a reasonable explanation for deep reinforcement learning models. The limitations of existing algorithms cannot well meet the needs of practical applications for interpretable algorithms.

[0010] The saliency map algorithm (Saliencymap) that interprets the perturbation input needs to perform regional Gaussian blurring on the input image at a certain interval. Each time the blur is performed, the blurred image is input into the network, and the difference between the obtained value and the value obtained when the original image enters the network is taken to obtain the degree of influence of the region on the model decision. In this way, it is not easy to obtain the influence of specific features in the image on the model decision through uniform blurring. When the blur range is small, it cannot cover the entire feature, and only the influence of a part of the feature on the decision can be obtained; when the blur range is relatively large, it is easy to cover multiple features, and the influence of a certain feature on the model decision cannot be obtained, which is not conducive to accurately analyzing the influence of each feature of the image on the model decision;

[0011] The LIME algorithm, which trains simple models to approximate complex models, uses a simple model to approximate a complex classification network. It uses a simple one-dimensional linear model to quantize and perturb the input image in one dimension, and then approximates the original model. Finally, the model can be interpreted by looking at the size of the coefficients of the linear model. This method can well explain the impact of the characteristics of the input image on the model's decision. However, LIME will only interpret one sample at a time, and a new model needs to be established each time. Although this algorithm is relatively general and accurate, it takes a long time to use, and it is difficult to use data to update the network. It is not very suitable for scenes with fast changes and high speed requirements. Summary of the invention

[0012] The embodiments of the present invention provide a method and system for the interpretability of a deep reinforcement learning model in an unmanned driving scenario, so as to at least solve the technical problem that the prior art cannot accurately analyze the impact of various features of an image on the model's decision-making.

[0013] According to an embodiment of the present invention, a method for interpreting a deep reinforcement learning model in an unmanned driving scenario is provided, comprising the following steps:

[0014] Select a suitable simulation environment and a suitable deep reinforcement learning algorithm, and obtain a converged reinforcement learning model through training;

[0015] The reinforcement learning model is input with pictures taken in unmanned driving scenarios, the pictures are divided into features and the influence of the features is quantitatively analyzed, the influence of each feature on the model decision is calculated, and the corresponding difference matrix is ​​obtained to obtain an improved network model.

[0016] Furthermore, the reinforcement learning model is input with pictures taken in the unmanned driving scene, the pictures are divided into features and the influence of the features is quantitatively analyzed, the influence of each feature on the model decision is calculated, and the corresponding difference matrix is ​​obtained. The improved network model includes:

[0017] First, the state image is obtained through the interaction between the model and the environment. The image is divided into a fixed number of blocks according to the features through superpixel segmentation. The irregular areas are blurred one by one by the Gaussian blur method to obtain an image set.

[0018] Then the image set and the original image are input into the network respectively, and the decision values ​​of the original image and the blurred image are obtained. The difference between the two is calculated to obtain the difference matrix.

[0019] The difference matrix is ​​upsampled so that the matrix size is equal to the size of the input image, and the value of the difference matrix is ​​multiplied by a preset multiple and superimposed on the original image.

[0020] Furthermore, A3C in deep reinforcement learning is selected as the algorithm for autonomous decision-making of intelligent agents in unmanned driving.

[0021] Furthermore, the unmanned driving environment selects the carla simulation environment, selects a suitable scene, and selects a picture as input.

[0022] Furthermore, before inputting the pictures taken in the unmanned driving scene into the reinforcement learning model, the method also includes: preprocessing the pictures taken in the unmanned driving scene.

[0023] Furthermore, preprocessing of images taken in the unmanned driving scenario includes:

[0024] Convert the input image into the form required for interpretation: find and segment the appropriate image features in the unmanned driving environment, and use the minimum number of segments to include the features required in the unmanned driving environment.

[0025] Furthermore, image segmentation combines adjacent pixels with similar texture, color, and brightness characteristics into visually meaningful irregular pixel blocks, and replaces a large number of pixels with a small number of pixels; image blurring is to take the average value of surrounding pixels for each pixel.

[0026] Furthermore, the saliency map algorithm is used to divide the features of the image and perform quantitative analysis of the influence of the features.

[0027] Furthermore, the method further comprises:

[0028] Present the explanation to the user in a form that is easy for the user to understand.

[0029] According to another embodiment of the present invention, a system for interpreting a deep reinforcement learning model in an unmanned driving scenario is provided, comprising:

[0030] The network model module is used to select a suitable simulation environment and a suitable deep reinforcement learning algorithm, and obtain a converged reinforcement learning model through training;

[0031] The explanatory algorithm module is used to input pictures taken in unmanned driving scenarios into the reinforcement learning model, divide the features of the pictures and conduct quantitative analysis of the influence of the features, calculate the influence of each feature on the model decision, and obtain the corresponding difference matrix to obtain an improved network model.

[0032] The interpretability method and system of the deep reinforcement learning model in the unmanned driving scenario in the embodiment of the present invention selects a suitable simulation environment and a suitable deep reinforcement learning algorithm, obtains a converged reinforcement learning model through training, inputs pictures taken in the unmanned driving scenario into the reinforcement learning model, divides the features of the pictures and performs quantitative analysis of the influence of the features, calculates the influence of each feature on the model decision, and obtains the corresponding difference matrix to obtain an improved network model. BRIEF DESCRIPTION OF THE DRAWINGS

[0033] The drawings described herein are used to provide a further understanding of the present invention and constitute a part of this application. The exemplary embodiments of the present invention and their descriptions are used to explain the present invention and do not constitute an improper limitation of the present invention. In the drawings:

[0034] Figure 1 The overall design framework diagram of the interpretability method and system of the deep reinforcement learning model in the unmanned driving scenario of the present invention;

[0035] Figure 2 This is a workflow diagram of the interpretability method and system of the deep reinforcement learning model in the unmanned driving scenario of the present invention. DETAILED DESCRIPTION

[0036] In order to enable those skilled in the art to better understand the scheme of the present invention, the technical scheme in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work should fall within the scope of protection of the present invention.

[0037] It should be noted that the terms "first", "second", etc. in the specification and claims of the present invention and the above-mentioned drawings are used to distinguish similar objects, and are not necessarily used to describe a specific order or sequence. It should be understood that the data used in this way can be interchanged where appropriate, so that the embodiments of the present invention described herein can be implemented in an order other than those illustrated or described herein. In addition, the terms "including" and "having" and any variations thereof are intended to cover non-exclusive inclusions, for example, a process, method, system, product or device that includes a series of steps or units is not necessarily limited to those steps or units that are clearly listed, but may include other steps or units that are not clearly listed or inherent to these processes, methods, products or devices.

[0038] The present invention is a novel XRL algorithm, which provides a fast and accurate solution for the interpretation and visualization of the decision of the deep reinforcement learning model; quantifies the influence of the features determined in the input image on the model decision; reduces the number of superpixel blocks in each image to increase the speed; and does not rely on a specific model in the design process to design a general interpretable algorithm that adapts to various practical scenarios. Therefore, the XRL proposed in the present invention is independent of a certain model (Model-free) and adapts to the actual scenario problems. XRL should also have certain flexibility and scalability so as to adapt to various scenarios with different feature numbers.

[0039] The problem to be solved by the present invention is to solve the unknowable problem of deep reinforcement learning model in unmanned driving scenarios and its extension problems by using algorithms such as superpixel segmentation, Gaussian blur, saliency map, and deep reinforcement learning, and to make users understand the favorable and unfavorable factors in the decision-making process through explanatory algorithms, and to present them to users through good human-computer interaction. The decision-making basis of the intelligent body is presented to users, increasing the user's trust in the unmanned driving model.

[0040] The basic contents of the technical solution of the present invention include the following:

[0041] 1. Design a general explanatory algorithm that is applicable to every series of deep reinforcement learning algorithms;

[0042] 2. Design of input image preprocessing in the interpretive algorithm process;

[0043] 3. Explain the impact of decision making through the network and preprocessed images;

[0044] 4. Present the explanation content to users through visualization technology.

[0045] The interpretability method and overall system design framework of the deep reinforcement learning model in the unmanned driving scenario consists of three parts: the network model part, the interpretable algorithm part, and the network improvement part, such as Figure 1 shown.

[0046] (1) Network model part

[0047] The network model part includes the selection of deep reinforcement learning algorithms, scenario design and model training in unmanned driving scenarios.

[0048] The present invention needs to select a suitable simulation environment and a suitable deep reinforcement learning algorithm in advance. By comparison, A3C (Asynchronous advantage actor-critic) in deep reinforcement learning is selected as the algorithm for autonomous decision-making of the intelligent agent in unmanned driving. The unmanned driving environment selects the carla simulation environment, selects a suitable scene, and selects a picture as input: then through training, a converged reinforcement learning model is finally obtained, at which point the model required by the present invention is obtained.

[0049] (2) Explanatory algorithm part

[0050] The explanatory algorithm part includes modules such as image preprocessing, saliency map operation (solving the difference matrix), and visualization.

[0051] Among them, the results of image preprocessing will be conducive to the operation of the saliency map algorithm (Saliencymap) module and the division of features, and will be conducive to the quantitative analysis of the influence of the explanatory algorithm on the features; the saliency map algorithm (Saliencymap) module will calculate the influence of each feature on the model decision and obtain the corresponding difference matrix, thereby obtaining the important factors in the decision-making process of the model; the visualization module presents the explanation content to the user in a form that is easy for the user to understand.

[0052] (3) Network improvement

[0053] By explaining the information, useful information is strengthened and unimportant information is isolated, making the network more effective and further verifying the effect of the explanation.

[0054] The interpretability methods and systems of deep reinforcement learning models in autonomous driving scenarios need to meet three basic requirements:

[0055] (1) Autonomous driving scenarios. The scenarios should be as rich as possible, close to the real situation, and a convergent deep reinforcement learning model should be obtained.

[0056] (2) Explanatory algorithm: The preprocessing part tries to make the number of features extracted from the image as appropriate as possible.

[0057] (3) The area of ​​the saliency map should be as convergent as possible and not too dispersed.

[0058] Based on these three requirements, the present invention designs an interpretable method and system workflow of the deep reinforcement learning model in the unmanned driving scenario, such as Figure 2 shown.

[0059] After the present invention obtains the required model, it begins to interpret the model. First, the state image is obtained by the interaction between the model and the environment. The image is divided into a fixed number of blocks according to the features through superpixel segmentation, and then the image set is blurred in sequence by the Gaussian Blur method for irregular areas. Then, the image set and the original image are input into the network respectively, so that the decision values ​​of the original image and the blurred image are obtained, and the difference between the two is made to obtain a difference matrix. The difference matrix is ​​upsampled so that the matrix size is equal to the size of the input image, and the value of the difference matrix is ​​multiplied by a certain multiple and superimposed on the original image, so as to be presented to the user in the form of a saliency map. Then, the area of ​​the significant part of the saliency enhancement is obtained, and the purpose of improving the network model is achieved.

[0060] Among them, the image preprocessing problem of XRL is the premise of XRL analysis. Converting the input image into the form required for the interpretation of the present invention can be described as: finding and segmenting the appropriate image features in the unmanned driving environment, and not segmenting some unimportant or relatively small features, using the minimum number of segmentation blocks to include the features required in the unmanned driving environment, thereby greatly reducing the time consumed, thereby achieving the desired effect. The superpixel segmentation and Gaussian blur preprocessing algorithm used in the present invention well realizes the segmentation of the main features of the input image and the Gaussian blur of irregular features, and is a good preprocessing method that can be used for deep reinforcement learning model interpretation.

[0061] Image preprocessing follows the traditional image processing algorithm process: Image segmentation divides adjacent pixels with similar texture, color, brightness and other characteristics into irregular pixel blocks with certain visual significance, and replaces a large number of pixels with a small number of pixels. Image blur can be understood as taking the average value of surrounding pixels for each pixel.

[0062] The characteristic of this explanatory algorithm is that the importance of features is quantified in each explanation process, and there are two positive and negative decisions, respectively. Therefore, only the parts that are different from the traditional explanatory algorithm will be described:

[0063] Blurring irregular areas: Blurring the image eliminates some features so that the image is different from the original image in terms of features, which is convenient for subsequent comparison of the strategy obtained with the original image. Regarding the blurred area, it is important to note that the blurred area should transition smoothly with the unblurred area as much as possible to avoid the overly obvious boundary between the blurred and unblurred parts affecting the model's decision-making.

[0064] Feature interpretation: The image is processed according to the feature area, and the degree of influence of the feature area on the model decision is obtained and quantified. Through normalization processing, each feature is made comparative, and the positive and negative influence of each feature on the decision is calculated. The data obtained is conducive to the subsequent update of the model.

[0065] Improve the model based on the explanation: The explanation we get can tell us what are the favorable and unfavorable factors in the normal operation of the model. At the same time, when the model fails, we can also understand which specific feature of the input image caused the system decision failure. With this information, we can improve the model.

[0066] The key points and intended protection points of the present invention are at least:

[0067] 1. XRL overall design plan;

[0068] 2. Visualization of image preprocessing methods;

[0069] 3. Improvement algorithm of deep reinforcement learning model based on interpretability.

[0070] The present invention aims at the interpretability of deep reinforcement learning models in unmanned driving scenarios to solve the opacity of DRL models in such scenarios, explain their decisions, and provide a visual human-computer interaction interface. It explains the interpretability of deep reinforcement learning models to a certain extent, increases user trust, and provides a basis for model improvement. The present invention mainly embodies the following advantages:

[0071] Quantify and compare the impact of input image features on model decisions to obtain important features;

[0072] The explanation system is composed of modules, which has the characteristics of high flexibility and good scalability.

[0073] In the image preprocessing stage, the irregular regions of the image features can be smoothed and blurred so that the blurred regions smoothly intersect with the unblurred regions;

[0074] The deep reinforcement learning model can be further improved through the explained content, which is a part not covered by the current explanation system and model improvement system.

[0075] Using the unmanned driving environment as the experimental platform, the XRL and model improvement scheme were verified, the XRL algorithm was verified, and the algorithm was visualized through the simulation platform to enable users to understand the basis of the model decision-making and improve the model based on this basis.

[0076] The alternatives of the present invention are at least:

[0077] 1. The XRL system is scalable and can be combined with expansion modules to meet customer needs, such as adding or changing the image preprocessing process, changing the image perturbation method, changing the difference calculation method, etc.

[0078] 2. It is proposed to improve the model through interpretability, enhance positive features, suppress negative features, and achieve model improvement.

[0079] The serial numbers of the above embodiments of the present invention are only for description and do not represent the advantages or disadvantages of the embodiments.

[0080] In the above embodiments of the present invention, the description of each embodiment has its own emphasis. For parts that are not described in detail in a certain embodiment, reference can be made to the relevant descriptions of other embodiments.

[0081] In the several embodiments provided in this application, it should be understood that the disclosed technical content can be implemented in other ways. Among them, the system embodiments described above are only schematic. For example, the division of units can be a logical function division. There may be other division methods in actual implementation. For example, multiple units or components can be combined or integrated into another system, or some features can be ignored or not executed. Another point is that the mutual coupling or direct coupling or communication connection shown or discussed can be through some interfaces, indirect coupling or communication connection of units or modules, which can be electrical or other forms.

[0082] The units described as separate components may or may not be physically separated, and the components shown as units may or may not be physical units, that is, they may be located in one place or distributed over multiple units. Some or all of the units may be selected according to actual needs to achieve the purpose of the present embodiment.

[0083] In addition, each functional unit in each embodiment of the present invention may be integrated into one processing unit, or each unit may exist physically separately, or two or more units may be integrated into one unit. The above-mentioned integrated unit may be implemented in the form of hardware or in the form of software functional units.

[0084] If the integrated unit is implemented in the form of a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, in essence, or the part that contributes to the prior art, or all or part of the technical solution can be embodied in the form of a software product, and the computer software product is stored in a storage medium, including a number of instructions for a computer device (which can be a personal computer, a server or a network device, etc.) to perform all or part of the steps of the methods of each embodiment of the present invention. The aforementioned storage medium includes: U disk, read-only memory (ROM, Read-Only Memory), random access memory (RAM, Random Access Memory), mobile hard disk, magnetic disk or optical disk and other media that can store program codes.

[0085] The above is only a preferred embodiment of the present invention. It should be pointed out that for ordinary technicians in this technical field, several improvements and modifications can be made without departing from the principle of the present invention. These improvements and modifications should also be regarded as the scope of protection of the present invention.

Claims

1. A method for interpreting deep reinforcement learning models in unmanned driving scenarios, characterized in that: The following steps are involved: Select a suitable simulation environment and a suitable deep reinforcement learning algorithm, and obtain a converged reinforcement learning model through training; The reinforcement learning model is fed with pictures taken in unmanned driving scenarios, and the pictures are divided into features and the influence of the features is quantitatively analyzed. The influence of each feature on the model decision is calculated, and the corresponding difference matrix is ​​obtained to obtain an improved network model. The specific steps include: First, the state image is obtained through the interaction between the model and the environment. The image is divided into a fixed number of blocks according to the features through superpixel segmentation. The irregular areas are blurred one by one by the Gaussian blur method to obtain an image set. Then the image set and the original image are input into the network respectively, and the decision values ​​of the original image and the blurred image are obtained. The difference between the two is calculated to obtain the difference matrix. The difference matrix is ​​upsampled so that the matrix size is equal to the size of the input image, and the value of the difference matrix is ​​multiplied by a preset multiple and superimposed on the original image.

2. The interpretability method of the deep reinforcement learning model in the unmanned driving scenario according to claim 1 is characterized in that: A3C in deep reinforcement learning is selected as the algorithm for autonomous decision-making of intelligent agents in unmanned driving.

3. The interpretability method of the deep reinforcement learning model in the unmanned driving scenario according to claim 1 is characterized in that: For the unmanned driving environment, select the carla simulation environment, select the appropriate scene, and select the picture as input.

4. The interpretability method of the deep reinforcement learning model in the unmanned driving scenario according to claim 1, characterized in that: Before inputting the pictures taken in the unmanned driving scene into the reinforcement learning model, the method also includes: preprocessing the pictures taken in the unmanned driving scene.

5. The interpretability method of the deep reinforcement learning model in the unmanned driving scenario according to claim 4 is characterized in that: Preprocessing of images taken in unmanned driving scenarios includes: Convert the input image into the form required for interpretation: find and segment the appropriate image features in the unmanned driving environment, and use the minimum number of segments to include the features required in the unmanned driving environment.

6. The interpretability method of the deep reinforcement learning model in the unmanned driving scenario according to claim 5 is characterized in that: Image segmentation combines adjacent pixels with similar texture, color, and brightness features into visually meaningful irregular pixel blocks, and replaces a large number of pixels with a small number of pixels.

7. The interpretability method of the deep reinforcement learning model in the unmanned driving scenario according to claim 1, characterized in that: Use the saliency map algorithm to divide the features of the image and perform quantitative analysis of the influence of the features.

8. The interpretability method of the deep reinforcement learning model in the unmanned driving scenario according to claim 1, characterized in that: The method further comprises: Present the explanation to the user in a form that is easy for the user to understand.

9. An interpretability system for deep reinforcement learning models in unmanned driving scenarios, characterized in that: include: The network model module is used to select a suitable simulation environment and a suitable deep reinforcement learning algorithm, and obtain a converged reinforcement learning model through training; The explanatory algorithm module is used to input the pictures taken in the unmanned driving scene into the reinforcement learning model, divide the features of the pictures and conduct quantitative analysis of the influence of the features, calculate the influence of each feature on the model decision, and obtain the corresponding difference matrix to obtain the improved network model; The explanatory algorithm module includes: First, the state image is obtained through the interaction between the model and the environment. The image is divided into a fixed number of blocks according to the features through superpixel segmentation. The irregular areas are blurred one by one by the Gaussian blur method to obtain an image set. Then the image set and the original image are input into the network respectively, and the decision values ​​of the original image and the blurred image are obtained. The difference between the two is calculated to obtain the difference matrix. The difference matrix is ​​upsampled so that the matrix size is equal to the size of the input image, and the value of the difference matrix is ​​multiplied by a preset multiple and superimposed on the original image.

Images