Debiased model attribution evaluation method and device based on random ablation

Abstract

The present application relates to a debiased model attribution evaluation method and device based on random ablation, wherein the method comprises: performing random ablation processing on the acquired original image based on the information-free value to obtain a data set; effectively alleviating the missing bias problem in the traditional feature ablation method; performing random ablation training on the pre-built deep learning model according to the data set and the preset loss function to obtain a target model that is robust to missing features; performing attribution evaluation on the test image based on the target model to obtain an attribution evaluation result; through the present application, the problem of the inability to balance the missing bias and the computational cost in the related art and the high computational cost of the evaluation is solved, and the scene of feature loss is simulated by randomly ablating features, so that the target model adapts to the feature loss, thereby effectively alleviating the missing bias problem in the evaluation stage, so as to achieve a balance between the missing bias and the computational cost and reduce the computational cost of the attribution evaluation.

Description

Technical Field

[0001] The present application relates to the field of image processing technology, and in particular to a method and device for evaluating debiased model attribution based on random ablation. Background Art

[0002] Traditional feature ablation methods measure the impact of features on model decisions by gradually removing them and observing changes in model performance. However, due to the inconsistent data distribution caused by feature removal, the evaluation results are affected by missingness bias, reducing the reliability of the evaluation. Therefore, methods such as the Reliable Interpretable Assessment (ROAR) algorithm are used to mitigate missingness bias by retraining the model, but this is computationally expensive and carries the risk of information leakage.

[0003] There is an inability to balance missing bias and computational cost in related technologies, and the computational cost of evaluation is high. No effective solution has been proposed so far. Summary of the Invention

[0004] In this embodiment, a debiased model attribution evaluation method and apparatus based on random ablation are provided to solve the problem in related technologies that the missing bias and computational cost cannot be balanced, resulting in high computational cost of evaluation.

[0005] First, in this embodiment, a debiased model attribution evaluation method based on random ablation is provided, including:

[0006] Based on the no-information value, the acquired original image is randomly ablated to obtain a data set;

[0007] According to the data set and the preset loss function, the pre-built deep learning model is trained with random ablation to obtain a target model;

[0008] Based on the target model, an attribution evaluation is performed on the test image to obtain an attribution evaluation result.

[0009] In some embodiments, random ablation is performed on the acquired original image based on the no-information value to obtain a data set, including:

[0010] Divide the acquired original image into several blocks;

[0011] Randomly removing a target number of the blocks in the original image;

[0012] Based on a mask operation, the removed blocks in the original image are replaced with the non-information values ​​to obtain the data set; the non-information values ​​are zero values ​​or mean values.

[0013] In some embodiments, randomly removing a target number of the blocks in the original image includes:

[0014] Based on the preset removal ratio, determine the target number of random removals;

[0015] Based on the target number, a target number of blocks in the original image are randomly removed.

[0016] In some embodiments, random ablation training is performed on a pre-built deep learning model based on the data set and a preset loss function to obtain a target model, including:

[0017] Constructing the deep learning model and setting the loss function of the deep learning model;

[0018] Training data is selected from the data set, the training data is randomly ablated and then input into the deep learning model, and iterative training is performed with minimization of the loss function as the optimization goal to obtain a target model.

[0019] In some embodiments, the loss function is a cross entropy loss function;

[0020] The gradient descent method is used to update the model parameters during the iterative training process of the deep learning model.

[0021] In some embodiments, performing attribution evaluation on a test image based on the target model to obtain an attribution evaluation result includes:

[0022] Determine the attribution value of the test image;

[0023] Based on the attribution value and the feature removal strategy, removing features from the test image to obtain an ablation image;

[0024] Inputting the ablation image into the target model to record the performance change of the target model and obtain a performance attenuation curve;

[0025] The performance degradation curve is analyzed to obtain an attribution evaluation result.

[0026] In some embodiments, the method further comprises:

[0027] Inputting the medical image to be identified into the target model for identification to obtain an identification result;

[0028] An attribution method is used to perform attribution analysis on the medical image to obtain an attribution result.

[0029] In a second aspect, a random ablation-based debiasing model attribution evaluation device is provided in this embodiment, comprising: a processing module, a training module, and an attribution evaluation module;

[0030] The processing module is used to perform random ablation processing on the acquired original image based on the no-information value to obtain a data set;

[0031] The training module is used to perform random ablation training on the pre-built deep learning model according to the data set and the preset loss function to obtain a target model;

[0032] The attribution evaluation module is used to perform attribution evaluation on the test image based on the target model to obtain an attribution evaluation result.

[0033] In a third aspect, a computer device is provided in this embodiment, comprising a memory, a processor, and a computer program stored on the memory and executable on the processor. When the processor executes the computer program, the random ablation-based debiased model attribution evaluation method described in the first aspect is implemented.

[0034] In a fourth aspect, a storage medium is provided in this embodiment, on which a computer program is stored. When the program is executed by a processor, the random ablation-based debiased model attribution evaluation method described in the first aspect is implemented.

[0035] Compared with the related art, the debiased model attribution evaluation method and device based on random ablation provided in this embodiment obtains a data set by performing random ablation processing on the acquired original image based on the information-free value; it can effectively alleviate the missing bias problem in the traditional feature ablation method; according to the data set and the preset loss function, the pre-built deep learning model is trained by random ablation to obtain a target model that is robust to missing features; based on the target model, the test image is attribution evaluated to obtain the attribution evaluation result, which solves the problem in the related art that the missing bias and the computational cost cannot be balanced, and the computational cost of the evaluation is high. By randomly ablating features to simulate the scenario of feature loss, the target model adapts to the feature loss, thereby effectively alleviating the missing bias problem in the evaluation stage, so as to achieve a balance between missing bias and computational cost and reduce the computational cost of attribution evaluation.

[0036] The details of one or more embodiments of the present application are set forth in the following drawings and description to make other features, objects, and advantages of the present application more readily apparent. BRIEF DESCRIPTION OF THE DRAWINGS

[0037] The drawings described herein are used to provide a further understanding of the present application and constitute a part of the present application. The illustrative embodiments of the present application and their descriptions are used to explain the present application and do not constitute an improper limitation on the present application. In the drawings:

[0038] Figure 1This is a hardware structure block diagram of a terminal device for a debiased model attribution evaluation method based on random ablation provided in one embodiment of the present application;

[0039] Figure 2 This is a flowchart of a debiased model attribution evaluation method based on random ablation provided in one embodiment of the present application;

[0040] Figure 3 is a flowchart of step S210;

[0041] Figure 4 is a flowchart of step S220;

[0042] Figure 5 is a flowchart of step S230;

[0043] Figure 6 This is a structural block diagram of a debiased model attribution evaluation device based on random ablation provided in one embodiment of the present application.

[0044] In the figure: 102, processor; 104, memory; 106, transmission device; 108, input and output device; 210, processing module; 220, training module; 230, attribution evaluation module. DETAILED DESCRIPTION

[0045] In order to more clearly understand the purpose, technical solutions and advantages of the present application, the present application is described and illustrated below in conjunction with the accompanying drawings and embodiments.

[0046] Unless otherwise defined, technical or scientific terms used in this application shall have the ordinary meanings as understood by persons of ordinary skill in the art to which this application belongs. The terms "a," "an," "the," "these," and similar expressions in this application do not denote limitations on quantity and may be singular or plural. The terms "comprise," "include," "have," and any variations thereof, as used in this application, are intended to cover non-exclusive inclusions. For example, a process, method, system, product, or device comprising a series of steps or modules (units) is not limited to the listed steps or modules (units) but may include unlisted steps or modules (units) or other steps or modules (units) inherent to the process, method, product, or device. The terms "connected," "connected," "coupled," and similar expressions used in this application are not limited to physical or mechanical connections but may include electrical connections, whether direct or indirect. As used in this application, "plurality" means two or more. "And / or" describes an association between associated objects, indicating that three possible relationships exist. For example, "A and / or B" can mean: A exists alone; A and B exist simultaneously; or B exists alone. Generally, the character " / " indicates that the objects in the preceding and following relationship are in an "or" relationship. The terms "first", "second", "third", etc. involved in this application are only used to distinguish similar objects and do not represent a specific ordering of the objects.

[0047] The method embodiment provided in this embodiment can be executed in a terminal, a computer or a similar computing device. For example, running on a terminal, Figure 1 : is a hardware structure block diagram of a terminal of the random ablation-based debiased model attribution evaluation method of this embodiment. Figure 1 As shown, the terminal may include one or more ( Figure 1 The processor 102 (only one is shown) and a memory 104 for storing data, wherein the processor 102 may include but is not limited to a processing device such as a microprocessor MCU or a programmable logic device FPGA. The terminal may also include a transmission device 106 for communication functions and an input / output device 108. It will be understood by those skilled in the art that Figure 1 The structure shown is only for illustration and does not limit the structure of the above terminal. Figure 1 More or fewer components than shown, or with Figure 1 Different configurations shown.

[0048] The memory 104 can be used to store computer programs, for example, software programs and modules of application software, such as the computer program corresponding to the debiased model attribution evaluation method based on random ablation in this embodiment. The processor 102 executes various functional applications and data processing by running the computer program stored in the memory 104, that is, to implement the above-mentioned method. The memory 104 may include a high-speed random access memory, and may also include a non-volatile memory, such as one or more magnetic storage devices, flash memory, or other non-volatile solid-state memory. In some instances, the memory 104 may further include a memory remotely located relative to the processor 102, and these remote memories may be connected to the terminal via a network. Examples of the above-mentioned network include, but are not limited to, the Internet, an intranet, a local area network, a mobile communication network, and combinations thereof.

[0049] Transmission device 106 is used to receive or transmit data via a network. This network may include a wireless network provided by the terminal's communications provider. In one embodiment, transmission device 106 includes a network interface controller (NIC), which can be connected to other network devices via a base station to enable communication with the Internet. In another embodiment, transmission device 106 may be a radio frequency (RF) module, which is used to communicate with the Internet wirelessly.

[0050] In this embodiment, a debiased model attribution evaluation method based on random ablation is provided. Figure 2 Flowchart of the random ablation-based debiased model attribution evaluation method of this embodiment is shown in FIG. Figure 2 As shown, the process includes the following steps:

[0051] Step S210, performing random ablation processing on the acquired original image based on the no-information value to obtain a data set;

[0052] Step S220: Perform random ablation training on the pre-built deep learning model according to the data set and the preset loss function to obtain a target model;

[0053] Step S230 : performing attribution evaluation on the test image based on the target model to obtain an attribution evaluation result.

[0054] Specifically, in actual application, the methods for obtaining original images and test images in the embodiments of the present application include but are not limited to obtaining original images and test images that meet the above requirements from pre-stored images in a database; it can also be downloading original images and test images that meet the requirements from a network platform; it can also be generating corresponding original images and test images according to needs, etc. The embodiments of the present application do not limit the method for obtaining original images and test images.

[0055] In this context, "no-information values" refer to pixel values ​​that cannot reflect the actual scene or measurement information. Therefore, performing random ablation on the acquired original image based on the no-information values ​​can be considered as randomly selecting features in the original image for ablation using methods such as zero-filling, masking, or intelligent interpolation filling based on the no-information values, thereby generating a dataset to simulate a scenario with feature loss. Preferably, the most important or least important features can be randomly selected for ablation, but this is not a limitation.

[0056] The deep learning model is pre-built, and its model architecture includes but is not limited to CNN architecture, ViT architecture, GAN architecture, etc. The loss function is pre-set, including but not limited to contrast loss function, KL divergence loss function, and cross entropy loss function. After the model architecture, data set, and loss function of the deep learning model are determined, the pre-built deep learning model can be trained with random ablation according to the data set and the preset loss function. That is, in each iterative training process, the most important or least important feature in the data set is randomly selected for ablation, and the loss function is used as the optimization target. When the deep learning model converges, a target model that is robust to missing features is obtained. Furthermore, the target model that is robust to missing features is applied to attribution evaluation. Based on the target model, the test image is subjected to attribution evaluation to obtain the attribution evaluation result, thereby effectively alleviating the missing bias problem in the evaluation stage and significantly reducing the computational cost.

[0057] In related technologies, the traditional feature ablation method is to measure the impact of features on model decisions by gradually removing features and observing changes in model performance; however, due to the inconsistent data distribution caused by feature removal, the evaluation results are affected by missing bias, which reduces the reliability of the evaluation. Therefore, methods such as the Reliable Explanation and Assessment (ROAR) algorithm are used to mitigate missing feature bias by retraining the model, but this approach is computationally expensive. This means that the related art cannot balance missing feature bias with computational cost, resulting in high computational cost for evaluation. In this embodiment, random ablation is performed on the acquired original image based on the information-free value to obtain a dataset, which effectively mitigates the missing feature bias problem in traditional feature ablation methods. Based on the dataset and a preset loss function, a pre-built deep learning model is trained with random ablation to obtain a target model that is robust to missing features. Based on the target model, attribution evaluation is performed on the test image to obtain an attribution evaluation result. This means that multiple attribution methods can be evaluated with only a single training run, significantly reducing computational cost and making them suitable for large-scale applications. Furthermore, since only a single training run is required, there is no need to retrain the model for each attribution method, reducing the risk of information leakage. This approach addresses the problem of high computational cost for evaluation and the inability to balance missing feature bias with computational cost in the related art. By randomly ablating features to simulate a scenario where feature loss occurs, the target model adapts to feature loss, effectively mitigating the missing feature bias during the evaluation phase, balancing missing feature bias with computational cost and reducing the computational cost of attribution evaluation.

[0058] The following is a detailed description of each of the above steps:

[0059] In some of these embodiments, Figure 3 As shown, the process of performing random ablation on the acquired original image based on the no-information value in step S210 to obtain a data set includes the following steps:

[0060] Step S211, dividing the acquired original image into several blocks;

[0061] Step S212, randomly removing a target number of blocks from the original image;

[0062] Step S213: Based on the mask operation, the removed blocks in the original image are replaced with non-information values ​​to obtain a data set; the non-information value is zero or the mean.

[0063] Specifically, in order to facilitate the processing of subsequent steps, the original image is first divided into several blocks with a block size of s×s; the number of blocks can be set by the usage scenario and is not limited. For example: the original image is X∈R H×W×C ; Where H is the image height, W is the image width, and C is the number of channels (e.g., C=3 for an RGB image). The set of all blocks of the original image can be expressed as: Among them, U i represents the i-th block, and m is the total number of blocks.

[0064] Then randomly remove the target number of blocks in the original image, and there is no limit on the number of removed blocks. Finally, replace the removed blocks in the original image with non-information values , and obtain a data set; among them, the non-information value is zero or the mean. This process can be implemented using masking operations, specifically:

[0065] The mask matrix is ​​represented as , where M(i, j) = 0 means the block at position (i, j) is removed, and M(i, j) = 1 means the block is retained.

[0066] The generation formula of the mask matrix M is:

[0067] ;

[0068] Then the image X in the dataset obtained after processing （t） Generated by masking:

[0069] ;

[0070] in, Indicates that the removed blocks are replaced with no information values ​​according to the mask matrix M .

[0071] Through this embodiment, the mask operation is used to quickly simulate the scenario of feature loss, and the subsequent model training process is adapted to reduce the subsequent calculation amount and improve training efficiency; and the removed features are replaced with non-information values ​​to ensure the consistency of the training and evaluation processes.

[0072] In some embodiments, the step S212 of randomly removing a target number of blocks from the original image includes the following steps:

[0073] Based on the preset removal ratio, determine the target number of random removals;

[0074] Based on the number of objects, randomly remove the number of blocks in the original image.

[0075] Specifically, the removal ratio can be expressed as , the number of blocks removed is .

[0076] Then from the set of all blocks of the original image Randomly select t blocks to remove, and the set of removed blocks is expressed as: ;in, represents random sorting, Indicates selecting the first t blocks.

[0077] Through this embodiment, the processing efficiency of random removal is improved, and the adjustment of random abnormal operation can be completed by adjusting the parameters of the above expression, which is convenient to use.

[0078] In some of these embodiments, Figure 4 As shown, in step S220, random ablation training is performed on the pre-built deep learning model according to the data set and the preset loss function to obtain the target model, including the following steps:

[0079] Step S221, constructing a deep learning model and setting a loss function of the deep learning model;

[0080] In step S222, training data is selected from the data set, the training data is randomly ablated and then input into the deep learning model, and iterative training is performed with the minimization of the loss function as the optimization goal to obtain the target model.

[0081] Specifically, a deep learning model is constructed using a model architecture from the CNN series, ViT series, or other series. An appropriate loss function is then selected, including contrastive loss, KL divergence loss, and cross-entropy loss. Since the dataset is already prepared, it is input as training data into the deep learning model to produce a target model that is robust to missing features. Preferably, the target model can be the RAINNet (Region-aware Adaptive Instance Normalization) model, which aims to address the image harmonization problem, i.e., eliminating inconsistencies in lighting, hue, and other aspects between the foreground and background in a stitched image.

[0082] The following takes the cross-entropy loss function as an example and uses the gradient descent method to update the model parameters of the deep learning model during iterative training of the deep learning model as an example to illustrate the training process. Specifically:

[0083] First, the model parameters Perform random initialization; secondly, the model training is performed in iterations. In each iteration, a small batch of data is randomly sampled from the dataset (the images in the dataset are randomly ablated so that there are differences in the datasets read at different times) and input into the deep learning model for calculation Finally, the model parameters are updated based on the calculation results. Through this process, the target model can recognize missing images, thus achieving the training goal.

[0084] During this iterative training process, the model parameters of the deep learning model are updated using the gradient descent method until the model converges to the target model. The model convergence condition is: minimizing the cross entropy loss function, that is, the cross entropy loss function is less than a preset threshold; or the number of iterations reaches a preset threshold.

[0085] Among them, the gradient descent method to update the model parameters can be expressed as:

[0086] ;

[0087] Where λ is the learning rate, represents the cross entropy loss function Gradients with respect to model parameters.

[0088] Cross Entropy Loss Function It can be expressed as:

[0089] ;

[0090] in, is the output of the target model.

[0091] In this embodiment, random ablation is used in training to enable the target model to adapt to feature-missing scenarios, that is, to maintain stable performance even when some features are removed, thereby optimizing the prediction performance of the target model in feature-missing scenarios.

[0092] Moreover, since the RAINNet model can maintain stable performance in the absence of features, it can provide more reliable and consistent evaluation results in attribution evaluation, thereby improving the reliability of the evaluation results.

[0093] In some of these embodiments, Figure 5 As shown, step S230 performs attribution evaluation on the test image based on the target model to obtain the attribution evaluation result, including the following steps:

[0094] Step S231, determining the attribution value of the test image;

[0095] Step S232: removing features from the test image based on the attribution value and the feature removal strategy to obtain an ablation image;

[0096] Step S233, inputting the ablation image into the target model to record the performance change of the target model and obtain a performance attenuation curve;

[0097] Step S234: Analyze the performance degradation curve to obtain an attribution evaluation result.

[0098] Specifically, after obtaining the target model, it can be used to perform attribution ablation testing to determine the attribution evaluation results, thereby evaluating the performance changes of each attribution method after removing important features. The target model can output the classification result of the input image; the attribution evaluation results can be the relationship between the classification result and the attribution values ​​of each feature (block or pixel) in the input image (the contribution of the feature (block or pixel) in the image to the target model), providing a reference for subsequent use and convenience for users.

[0099] First, calculate the attribution value. Specifically, you can use a variety of attribution methods (such as Integrated Gradients, Saliency, Guided Backpropagation, etc.) to calculate the attribution value of the test image. The attribution value represents the importance of each feature (pixel or block) in the image to the model prediction. If the attribution method is A, the attribution value is expressed as A(X)∈R H×W Preferably, after obtaining the attribution value, an attribution map can be generated based on the correspondence between the attribution value and the feature. The value of each pixel or block in the attribution map represents its contribution to the model prediction. The attribution map can then clearly provide feedback on the contribution.

[0100] Secondly, a feature ablation experiment is conducted, which includes removing features from the test image, recording the performance changes of the target model, and analyzing the performance degradation curve. The specific process is as follows:

[0101] 1. Feature removal of the test image: Sort all features in the test image according to the attribution value, and gradually remove the most important features (feature removal strategy of MoRF mode) or the least important features (feature removal strategy of LeRF mode) to obtain the ablation image. If the removal ratio is , the number of blocks removed is . Then the set of removed blocks is ;

[0102] in, Before selection The most important feature.

[0103] 2. Record the performance changes of the target model (performance evaluation): Input the ablation image into the target model, observe the changes in the target model performance, and obtain the performance decay curve.

[0104] 3. Analyze the performance decay curve: By recording the decay curve of the target model’s accuracy, the effectiveness of each attribution method can be evaluated. The accuracy after proportional features is , the performance degradation curve is Follow By analyzing the performance decay curve, we can measure the performance change of different attribution methods after removing important features. The faster the performance decay, the more accurately the attribution method can identify important features.

[0105] In other embodiments, the attribution methods may be ranked to select the most suitable attribution method for the target model.

[0106] Specifically: Determine the evaluation metric. Use the Area Difference (AD) metric to quantify the performance of the attribution method. The AD metric represents the difference between the performance decay curves in the MoRF mode and the LeRF mode. The larger the AD value, the more accurately the attribution method can distinguish important features from unimportant features. The expression of the AD metric can be:

[0107] ;

[0108] The attribution methods are then ranked based on the AD index. The higher the AD index value, the better the attribution method performance; it can achieve measurable performance and optimize user experience.

[0109] In some embodiments, the debiased model attribution evaluation method based on random ablation further includes the following steps:

[0110] Input the medical image to be identified into the target model for identification to obtain the identification result;

[0111] The attribution method is used to perform attribution analysis on medical images and obtain the attribution results.

[0112] This embodiment can be considered to be applied to the medical field and used to identify medical images to achieve pathology classification.

[0113] Medical films with missing parts due to various real-world factors (such as equipment artifacts and transmission compression) are input into the target model as the medical image to be identified. Specifically, the target model performs random ablation on the medical image to be identified, identifies the processed image, and outputs the recognition result. After the recognition result is obtained, attribution analysis is performed on the medical image using attribution methods to obtain the attribution result. The combined recognition and attribution results can assist users in further analysis and facilitate user experience.

[0114] In other embodiments, it can also be used in fields such as weather forecasting and traffic forecasting. The specific application implementation process can refer to the above embodiments and will not be repeated here.

[0115] The following uses different datasets as examples to illustrate the attribution evaluation of the RAINNet model.

[0116] Example 1 (attribution evaluation on the CIFAR-100 dataset):

[0117] 1. Random Ablation Training: Based on the original CIFAR-100 dataset, we divide the original 32×32 images into 1×1 blocks, randomly remove 25% of the blocks, and replace the removed blocks with zero values ​​as uninformative values ​​to generate a training dataset. The processed images are then fed into a ResNet-18 model (a deep learning model) for training, generating a RAINNet model as the target model.

[0118] 2. Attribution Ablation Test: Use the Integrated Gradients attribution method to calculate the attribution value of the processed image. Based on the RAINNet model, perform feature ablation experiments on the attribution results to evaluate the performance changes of the Integrated Gradients method after removing important features, and generate a performance degradation curve. By analyzing the performance degradation curve, the effectiveness of the Integrated Gradients attribution method is measured. This method can also be used to test other attribution methods and rank them.

[0119] Example 2 (attribution evaluation on the ImageNet-50 dataset):

[0120] 1. Random Ablation Training: Based on the original ImageNet-50 dataset, we divide the original 224×224 images into 14×14 blocks, randomly remove 50% of the blocks, and replace the removed blocks with their mean values ​​as uninformative values ​​to generate a training dataset. The processed images are then fed into a ResNet-18 model (a deep learning model) for training, generating a RAINNet model as the target model.

[0121] 2. Attribution Ablation Test: Use the Saliency method to calculate the attribution value of the processed image. A feature ablation experiment is performed on the attribution results based on the RAINNet model to evaluate the performance change of the Saliency method after removing important features. This performance degradation curve is generated. By analyzing the performance degradation curve, the effectiveness of the Saliency method is measured. This method can be used to test other attribution methods and rank them.

[0122] It should be noted that the steps shown in the above process or the flowchart in the accompanying drawings can be executed in a computer system such as a set of computer-executable instructions, and although a logical order is shown in the flowchart, in some cases, the steps shown or described can be executed in an order different from that shown here.

[0123] In this embodiment, a debiased model attribution evaluation device based on random ablation is also provided, which is used to implement the above-mentioned embodiments and preferred embodiments, and will not be repeated hereafter. The terms "module", "unit", "sub-unit", etc. used below may be a combination of software and / or hardware that implements predetermined functions. Although the devices described in the following embodiments are preferably implemented in software, implementation in hardware, or a combination of software and hardware, is also possible and conceivable.

[0124] Figure 6 is a structural block diagram of the random ablation-based debiased model attribution evaluation device of this embodiment, such as Figure 6 As shown, the apparatus includes: a processing module 210, a training module 220, and an attribution evaluation module 230;

[0125] A processing module 210 is configured to perform random ablation processing on the acquired original image based on the no-information value to obtain a data set;

[0126] The training module 220 is used to perform random ablation training on the pre-built deep learning model according to the data set and the preset loss function to obtain the target model;

[0127] The attribution evaluation module 230 is used to perform attribution evaluation on the test image based on the target model to obtain an attribution evaluation result.

[0128] Through the above-mentioned device, the problem of the inability to balance missing bias and computational cost in related technologies and the high computational cost of evaluation is solved. By randomly ablating features to simulate the scenario of feature loss, the target model is made to adapt to feature loss, thereby effectively alleviating the problem of missing bias in the evaluation stage, so as to achieve a balance between missing bias and computational cost and reduce the computational cost of attribution evaluation.

[0129] In some of the embodiments, the processing module 210 is further configured to divide the acquired original image into a plurality of blocks;

[0130] Randomly remove the target number of blocks in the original image;

[0131] Based on the mask operation, the removed blocks in the original image are replaced with uninformative values ​​to obtain a data set; the uninformative values ​​are zero values ​​or the mean.

[0132] In some embodiments, the processing module 210 is further configured to determine a target number of random removals based on a preset removal ratio;

[0133] Based on the number of objects, randomly remove the number of blocks in the original image.

[0134] In some embodiments, the training module 220 is further used to construct a deep learning model and set a loss function of the deep learning model;

[0135] Training data is selected from the dataset, randomly ablated, and then input into the deep learning model. It is then iteratively trained with the loss function minimization as the optimization goal to obtain the target model.

[0136] In some of the embodiments, the loss function is a cross entropy loss function;

[0137] The gradient descent method is used to update the model parameters during the iterative training of the deep learning model.

[0138] In some of these embodiments, the attribution evaluation module 230 is further configured to determine an attribution value for the test image;

[0139] Based on the attribution value and feature removal strategy, the features of the test image are removed to obtain the ablation image;

[0140] Input the ablation image into the target model to record the performance change of the target model and obtain the performance decay curve;

[0141] The performance degradation curve is analyzed to obtain the attribution evaluation results.

[0142] In some of the embodiments, the debiased model attribution evaluation device based on random ablation further includes: an application module;

[0143] The application module is used to input the medical image to be identified into the target model for identification and obtain the identification result;

[0144] The attribution method is used to perform attribution analysis on medical images and obtain the attribution results.

[0145] It should be noted that the above modules can be functional modules or program modules, and can be implemented through software or hardware. For modules implemented through hardware, the above modules can be located in the same processor; or the above modules can be located in different processors in any combination.

[0146] This embodiment further provides a computer device, including a memory and a processor, wherein the memory stores a computer program, and the processor is configured to run the computer program to execute the steps in any one of the above method embodiments.

[0147] Optionally, the computer device may further include a transmission device and an input / output device, wherein the transmission device is connected to the processor, and the input / output device is connected to the processor.

[0148] Optionally, in this embodiment, the processor may be configured to execute the following steps through a computer program:

[0149] S1, based on the no-information value, performs random ablation processing on the acquired original image to obtain a data set;

[0150] S2, based on the dataset and the preset loss function, performs random ablation training on the pre-built deep learning model to obtain the target model;

[0151] S3, based on the target model, performs attribution evaluation on the test image to obtain the attribution evaluation result.

[0152] It should be noted that, for specific examples in this embodiment, reference may be made to the examples described in the above embodiments and optional implementation modes, and will not be repeated in this embodiment.

[0153] In addition, in conjunction with the random ablation-based debiased model attribution evaluation method provided in the above embodiments, a storage medium may also be provided in this embodiment for implementation. The storage medium stores a computer program; when the computer program is executed by a processor, it implements any of the random ablation-based debiased model attribution evaluation methods in the above embodiments.

[0154] It should be noted that the information and data involved in this application are all authorized by the user or fully authorized by all parties and are legal for use.

[0155] It should be understood that the specific embodiments described herein are only used to explain this application and are not used to limit it. Based on the embodiments provided in this application, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of this application.

[0156] Obviously, the accompanying drawings are merely examples or embodiments of the present application. A person skilled in the art can also apply the present application to other similar situations based on these drawings without inventive effort. Furthermore, it is understandable that, although the work involved in this development process may be complex and lengthy, certain design, manufacturing, or production changes based on the technical content disclosed in this application are merely routine technical means for a person skilled in the art and should not be considered to constitute a deficiency in the disclosure of the present application.

[0157] The term "embodiment" as used in this application refers to specific features, structures, or characteristics described in conjunction with the embodiment that can be included in at least one embodiment of this application. The appearance of this phrase in various places in the specification does not necessarily mean that the embodiment is the same, nor does it mean that it is mutually exclusive with other embodiments and is independent or optional. It is understood, either explicitly or implicitly, by those skilled in the art that the embodiments described in this application can be combined with other embodiments when there is no conflict.

[0158] The above-described embodiments merely represent several implementation methods of the present application. While the descriptions are relatively specific and detailed, they should not be construed as limiting the scope of patent protection. It should be noted that a person of ordinary skill in the art may make various modifications and improvements without departing from the spirit of the present application, and these modifications and improvements fall within the scope of protection of the present application. Therefore, the scope of protection of the present application shall be determined by the appended claims.

Claims

1. A debiased model attribution evaluation method based on random ablation, characterized by: include: Based on the no-information value, the acquired original image is randomly ablated to obtain a data set, which includes: Divide the acquired original image into several blocks; Randomly removing a target number of the blocks in the original image; Based on a mask operation, the removed blocks in the original image are replaced with the non-information values ​​to obtain the data set; the non-information values ​​are zero or the mean; According to the data set and the preset loss function, the pre-built deep learning model is trained with random ablation to obtain a target model; Based on the target model, an attribution evaluation is performed on the test image to obtain an attribution evaluation result.

2. The debiased model attribution evaluation method based on random ablation according to claim 1, characterized in that: Randomly removing a target number of blocks from the original image, comprising: Based on the preset removal ratio, determine the target number of random removals; Based on the target number, a target number of blocks in the original image are randomly removed.

3. The debiased model attribution evaluation method based on random ablation according to any one of claims 1 to 2, characterized in that: According to the data set and the preset loss function, the pre-built deep learning model is trained with random ablation to obtain the target model, including: Constructing the deep learning model and setting the loss function of the deep learning model; Training data is selected from the data set, the training data is randomly ablated and then input into the deep learning model, and iterative training is performed with minimization of the loss function as the optimization goal to obtain a target model.

4. The debiased model attribution evaluation method based on random ablation according to claim 3, characterized in that: The loss function is a cross entropy loss function; The gradient descent method is used to update the model parameters during the iterative training process of the deep learning model.

5. The debiased model attribution evaluation method based on random ablation according to claim 1, characterized in that: Based on the target model, an attribution evaluation is performed on the test image to obtain an attribution evaluation result, including: Determine the attribution value of the test image; Based on the attribution value and the feature removal strategy, removing features from the test image to obtain an ablation image; Inputting the ablation image into the target model to record the performance change of the target model and obtain a performance attenuation curve; The performance degradation curve is analyzed to obtain an attribution evaluation result.

6. The debiased model attribution evaluation method based on random ablation according to claim 1, characterized in that: The method further comprises: Inputting the medical image to be identified into the target model for identification to obtain an identification result; An attribution method is used to perform attribution analysis on the medical image to obtain an attribution result.

7. A debiased model attribution evaluation device based on random ablation, characterized in that: include: Processing module, training module and attribution evaluation module; The processing module is used to perform random ablation processing on the acquired original image based on the no-information value to obtain a data set, which includes: Divide the acquired original image into several blocks; Randomly removing a target number of the blocks in the original image; Based on a mask operation, the removed blocks in the original image are replaced with the non-information values ​​to obtain the data set; the non-information values ​​are zero or the mean; The training module is used to perform random ablation training on the pre-built deep learning model according to the data set and the preset loss function to obtain a target model; The attribution evaluation module is used to perform attribution evaluation on the test image based on the target model to obtain an attribution evaluation result.

8. A computer device comprising a memory and a processor, characterized in that: A computer program is stored in the memory, and the processor is configured to run the computer program to perform the steps of the random ablation-based debiased model attribution evaluation method according to any one of claims 1 to 6.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the steps of the random ablation-based debiased model attribution evaluation method according to any one of claims 1 to 6 are implemented.

Images