An infrared small target detection method based on background probability modulation mixed experts

Abstract

The application discloses an infrared small target detection method based on background probability modulation mixed experts, comprising the following steps: 1. acquiring an infrared image and a corresponding target mask label, and performing data preprocessing to generate a batch of training samples, and defining a typical background category set; 2. constructing an infrared small target detection network, comprising: a CLIP text encoder, a CLIP image encoder, an infrared visual encoder and a modulated infrared small target detector, which are used for target detection on the batch of samples; 3. designing a loss function, comprising: a distillation loss function and a mask loss function, and training the infrared small target detection network; 4. using the trained infrared small target detection model to realize infrared small target detection. The application introduces a CLIP model to evaluate the background probability of the image to be detected, and dynamically associates the background probability predicted by the CLIP with the infrared visual representation by using a mixed expert model, so that effective infrared small target and background decoupling is realized.

Description

Technical Field

[0001] This invention belongs to the fields of deep learning and infrared image processing technology, specifically involving an infrared small target detection method based on background probability modulation hybrid expert, which can effectively perceive distant small targets by combining the background information of the whole image. Background Technology

[0002] Infrared small target detection aims to accurately segment faint, obscured targets from complex backgrounds, playing a crucial role in low-altitude surveillance, autonomous navigation, and security early warning. Despite its wide range of applications, this task is extremely challenging due to the weak difference in thermal radiation between the target and the surrounding background, the extremely long imaging distance, and the diversity of background clutter.

[0003] The rapid development of deep learning has driven the continuous advancement of data-driven infrared small target detection methods. By learning discriminative target-background representations from large-scale datasets, these methods significantly reduce their reliance on manually designed priors. Representative architectures include convolutional neural networks, visual Transformers, state-space models, and KAN networks. Although these methods have made significant progress in local feature extraction, global dependency modeling, computational efficiency, and nonlinear representation capabilities, the aforementioned pure vision models are still fundamentally limited by the lack of inherent information in images.

[0004] In recent years, the development of visual-language models has opened up new research directions for infrared small target detection through complementary image-text representation learning. Despite their effectiveness, existing visual-language model-based infrared small target detection methods still suffer from three limitations: First, the reliance on paired text descriptions hinders effective textless inference. Existing methods typically rely on image-text pairs, where the text descriptions are either generated by the visual-language model or manually designed. During the inference phase, such paradigms either require additional text input or experience performance degradation after removing the text input. In fact, paired text descriptions are not always available in real-world applications, thus limiting the practical application value of these methods. Second, insufficient CLIP knowledge transfer limits the infrared representation learning capability. Since CLIP is pre-trained on visible light images, effectively transferring its knowledge to the infrared domain is crucial for infrared small target detection. Existing strategies mainly focus on explicit cue engineering, and still have significant limitations in terms of language knowledge transfer. Furthermore, freezing or fine-tuning the CLIP visual encoder is insufficient to fully adapt to the characteristics of infrared images. Third, static feature fusion hinders the full utilization of CLIP-derived semantic information. Existing methods typically inject CLIP features into the detection model through attention-based static fusion mechanisms. However, the semantic information derived from CLIP varies significantly across different infrared images. Static fusion methods struggle to leverage these inter-sample semantic differences to dynamically adjust the detection model, limiting the model's robust target clutter decoupling capabilities in diverse contexts. Summary of the Invention

[0005] To overcome the problems of strong text dependence, insufficient CLIP knowledge transfer ability, and insufficient utilization of CLIP-derived semantic information in existing infrared small target detection methods based on visual language models, this invention proposes an infrared small target detection method based on background probability modulation hybrid expert. The aim is to achieve efficient reasoning without text dependence while learning natural language knowledge, and to realize dynamic and differentiated modulation of the detection model by visual language knowledge, thereby improving the robustness of infrared small target detection in various complex scenarios.

[0006] To achieve the above-mentioned objectives, the present invention adopts the following technical solution: The infrared small target detection method based on background probability modulation hybrid expert of the present invention is characterized by the following steps: Step 1: Acquire and preprocess infrared small target images with labeled information to obtain preprocessed infrared image samples. ,make The true value mask label is ,in, E, H, W These represent the number of channels, height, and width of the infrared image sample, respectively. Define a typical infrared background category set ;in, Indicates the first A typical infrared background category; K The number of typical infrared background categories; Step 2, construct an infrared small target detection network, including: CLIP text encoder CLIP image encoder Infrared visual encoder Modulated infrared small target detector, and for and Processing is performed to obtain Predicted mask results ;in, It uses the original text encoding structure that comes with the CLIP model. It adopts the standard ViT-B / 16 architecture. The ViT-Tiny / 16 structure is adopted. The modulated infrared small target detector consists of a coding layer, a background probability modulation hybrid expert layer, and a decoding layer. The coding layer consists of convolution and upsampling structures, and the decoding layer consists of convolution and upsampling structures. Step 3, based on and The final loss of constructing an infrared small target detection network ; Step 4, Freeze and Then, the gradient descent method is used to backpropagate the infrared small target detection network, and the final loss is calculated. The network parameters are optimized to obtain a trained infrared small target detection model; the trained text is then input. Input to The optimal text features are obtained through processing. And store offline; During the inference phase, the trained infrared visual encoder, the trained modulated infrared small target detector, and pre-stored... Small target detection is performed on the infrared image to be detected, and 0.5 is used as the segmentation threshold to obtain the optimal prediction mask result.

[0007] The infrared small target detection method based on background probability modulation hybrid expert described in this invention is also characterized in that step 2 includes: Step 2.1, for Perform tokenization encoding to obtain the encoded category token. ,in, Indicates the number of category tokens. Indicates the dimension of the token; Concatenate the two tokens to be learned to obtain the text input. ,in, Represents a prefix token. Represents a suffix token. Represents channel splicing. n represents the number of prefix tokens or suffix tokens; Step 2.2, will Input to CLIP text encoder After processing, text features are obtained. ; Step 2.3, will Input CLIP image encoder The process is performed to obtain the teacher intermediate features output from the 4th layer. The intermediate features of teachers output from layer 8 The intermediate features of teachers output from layer 12 And the teacher's global token extracted at the last layer. and will Input to linear mapping layer The image features of the teacher are then processed to obtain their characteristics. ; Step 2.4, Input infrared vision encoder The process is performed to obtain the student intermediate features output from the 4th layer. The intermediate features of students output from layer 8 The intermediate features of students output from layer 12 And the student global token extracted at the last level. ; and will Input to linear mapping layer The student image features are then processed to obtain the student image features. ; Step 2.5, , respectively with Perform cosine similarity calculation to obtain the teacher's logical value. Logical value of students And use the Softmax activation function to Processing is performed to obtain the background probability. ; Step 2.6, The input is processed by the modulated infrared small target detector to obtain... Predicted mask results .

[0008] Furthermore, step 2.6 includes: Step 2.6.1, through the coding layer Feature extraction is performed to obtain deep visual tokens. , where m represents the number of deep visual tokens; Step 2.6.2, will and The inputs are fed into the background probability modulation mixing expert layer for processing to obtain the modulation token. ; Step 2.6.3, will After being processed in the decoding layer, the input undergoes further processing using the Sigmoid activation function to obtain... Predicted mask results .

[0009] Furthermore, the background probability modulation hybrid expert layer in step 2.6.2 includes: Step a, first use the broadcast mechanism to... and The images are stitched together to obtain a visual representation with background probability awareness. ;right After normalization, the input is fed into the key mapping layer. The routing key matrix is ​​obtained through processing. ; Step b: Initialize a set of slot query matrices to be learned. Where e and s represent the number of experts and the number of slots for each expert, respectively; Step c, will Input to probability mapping layer The shape transformation is performed to obtain the background semantic query matrix. ; thereby and After concatenation, the final routing query matrix is ​​obtained. ; Step d, based on and The hyperspherical matched logic value is calculated using equation (5). : × (5) In equation (6), represent Norm, Represents the transpose operation; Step e, perform A along dimension m. After activation, the distribution weight of A is obtained. ; Step f, utilizing distribution weights right Perform weighted aggregation to obtain slot features. ; Step g: According to the number of slots s, sort the slot features Divided into e dedicated slot features ,in, Indicates the first Each unique slot feature; Will Enter to the number Expert Module In this way, the first equation can be obtained using equation (7). Enhanced exclusive slot features : (6) In equation (6), It is the first j A linear mapping layer with doubled dimensions. express Activation function It is a linear mapping layer that compresses the j-th dimension; Step h, along the number of slots s After splicing, the enhanced slot feature is obtained. ; Step i, along ( Dimension A After activation, the aggregate weight of A is obtained. Thus using right Weighted aggregation yields .

[0010] Furthermore, step 3 includes: Step 3.1, construct the segmentation loss using equation (1) : (1) In equation (1), This represents the binary cross-entropy loss function; Step 3.2, construct the characteristic distillation loss using equation (2). : (2) In equation (2), Indicates the first The intermediate features of teachers output by the layer. Indicates the first Student intermediate features output by layer ; Represents the mean squared error loss function. Indicates the first A multi-level linear mapping layer that aligns the student intermediate features and teacher intermediate features output by dimension. This represents a token linear mapping layer that aligns the student global token with the teacher global token in terms of dimensions. Step 3.3, construct the response-level distillation loss using equation (3). : (3) In equation (3), This is the distillation temperature coefficient. Let KL divergence loss function be used. Step 3.4, construct the final loss using equation (4). : (4).

[0011] The present invention provides an electronic device, including a memory and a processor, characterized in that the memory is used to store a program supporting the processor in performing the method described therein, and the processor is configured to execute the program stored in the memory.

[0012] The present invention discloses a computer-readable storage medium storing a computer program, characterized in that the computer program is executed by a processor to perform the steps of the method described thereon.

[0013] Compared with the prior art, the beneficial effects of the present invention are as follows: 1. This invention redefines the visual language model as a zero-shot background classifier. By predefining common background categories and calculating the background probability corresponding to the input image, image-level semantic priors are obtained. Thus, free text reasoning can be achieved without relying on additional text input such as target category or scene description. This solves the problem of dependence on paired text descriptions in existing visual language infrared small target detection methods and improves the practical application capability of the method.

[0014] 2. This invention proposes a collaborative knowledge transfer method, which achieves effective transfer of visual language model knowledge to the infrared domain through learnable text prompts and multi-angle visual knowledge distillation. By optimizing the background category text representation and using a visual teacher network to guide a lightweight student network to learn infrared visual features, this method solves the problems of insufficient visual language model knowledge transfer and difficulty in adapting to the characteristics of infrared imaging in existing methods. It improves the accuracy of background probability prediction and the feature representation capability of infrared images, while reducing inference computation overhead.

[0015] 3. This invention designs a background probability modulation hybrid expert layer, which uses background probability to dynamically control expert routing and adaptively modulate visual features. By establishing a dynamic correlation between background semantics and visual features in the hyperspherical embedding space, different experts can perform differentiated feature modeling for semantic changes in different scenarios. This solves the problems of static fusion of CLIP semantics and difficulty in utilizing semantic differences between samples in existing methods, effectively enhancing the decoupling ability between the target and complex background clutter, and improving the detection accuracy and robustness of small targets in diverse infrared scenarios. Attached Figure Description

[0016] Figure 1 This is a schematic diagram of the infrared small target detection method based on background probability modulation hybrid expert of the present invention; Figure 2 This is a schematic diagram of the background probability modulation hybrid expert layer of the present invention; Figure 3 This diagram illustrates how the knowledge transfer method of the present invention improves the number of model parameters, computational complexity, and performance. Figure 4 This is a schematic diagram illustrating the expert specialization effect of the background probability modulation hybrid expert layer of the present invention on three datasets; Figure 5 This is a diagram showing the detection results of small infrared targets according to the present invention. Detailed Implementation

[0017] In this embodiment, an infrared small target detection method based on background probability modulation hybrid expert includes the following steps: Step 1: Acquire and preprocess infrared small target images with labeled information to obtain preprocessed infrared image samples. ,make The true value mask label is ,in, E, H, W These represent the number of channels, height, and width of the infrared image sample, respectively.

[0018] Define a typical infrared background category set ;in, Indicates the first A typical infrared background category; K This represents the number of typical infrared background categories.

[0019] Step 2: Construct an infrared small target detection network based on background probability modulation hybrid expert, the network architecture of which is as follows: Figure 1 As shown, it includes: CLIP text encoder CLIP image encoder Infrared visual encoder Modulated infrared small target detector; for input and Processing is performed to obtain Predicted mask results ;in, It uses the original text encoding structure that comes with the CLIP model. It adopts the standard ViT-B / 16 architecture. It is a teacher model in the process of visual knowledge transfer. It adopts the ViT-Tiny / 16 structure. It is a student model in the process of visual knowledge transfer. The modulated infrared small target detector consists of an encoding layer, a background probability modulation hybrid expert layer, and a decoding layer. The encoding layer consists of convolution and upsampling structures, and the decoding layer consists of convolution and upsampling structures.

[0020] Step 2.1, to effectively transfer CLIP knowledge, such as Figure 1 As shown, a knowledge transfer method is proposed that combines learnable text prompts with a multi-view visual knowledge distillation strategy, and its effect is as follows: Figure 3 As shown, by using the knowledge transfer method, this network significantly reduces the number of model parameters and computational complexity, thereby reducing the dependence on hardware resources; at the same time, the pixel-level prediction accuracy (intersection over union ratio) and target-level prediction accuracy (detection probability and false alarm rate) of this network are significantly improved. For text branches, Perform tokenization encoding to obtain the encoded category token. , Indicates the number of category tokens. Indicates the dimension of the token; Concatenate the two tokens to be learned to obtain the text input. ,in, Represents a prefix token. Represents a suffix token. Represents channel splicing. n represents the number of prefix and suffix tokens, which needs to be learned. and Able to The text descriptions are optimized to make the text features more accurately aligned with the semantics of the infrared scene.

[0021] Step 2.2, will Input to CLIP text encoder After processing, text features are obtained. ; Step 2.3, for the visual branch, will Input CLIP image encoder The process is performed to obtain the teacher intermediate features output from the 4th layer. The intermediate features of teachers output from layer 8 The intermediate features of teachers output from layer 12 And the teacher's global token extracted at the last layer. and will Input to linear mapping layer The image features of the teacher are then processed to obtain their characteristics. .

[0022] Step 2.4, Input infrared vision encoder The process is performed to obtain the student intermediate features output from the 4th layer. The intermediate features of students output from layer 8 The intermediate features of students output from layer 12 And the student global token extracted at the last level. ; and will Input to linear mapping layer The student image features are then processed to obtain the student image features. .

[0023] Step 2.5, , respectively with Perform cosine similarity calculation to obtain the teacher's logical value. Logical value of students And use the Softmax activation function to Processing is performed to obtain the background probability. By pre-constructing common background categories in infrared scenes and calculating the matching probability between infrared images and each background category. This replaces the text input required by existing visual language infrared small target detection methods, thereby achieving an efficient paradigm that requires no additional text input.

[0024] Step 2.6, The input is processed by the modulated infrared small target detector to obtain... Predicted mask results ; Step 2.6.1, through the coding layer Feature extraction is performed to obtain deep visual tokens. , where m represents the number of deep visual tokens.

[0025] Step 2.6.2, will and The inputs are fed into the background probability modulation mixing expert layer for processing to obtain the modulation token. .

[0026] Step a, the structure of the background probability modulation hybrid expert layer is as follows: Figure 2 As shown, it calculates the hyperspherical matching logic value by modeling the correlation between background probability and visual features, thereby dynamically assigning visual tokens to experts. First, a broadcast mechanism is used to... and The images are stitched together to obtain a visual representation with background probability awareness. ;right After normalization, the input is fed into the key mapping layer. The routing key matrix is ​​obtained through processing. .

[0027] Step b: Initialize a set of slot query matrices to be learned. Where e and s represent the number of experts and the number of slots for each expert, respectively; Step c, will Input to probability mapping layer The shape transformation is performed to obtain the background semantic query matrix. ; thereby and After concatenation, the final routing query matrix is ​​obtained. .

[0028] Step d, based on and The hyperspherical matched logic value is calculated using equation (5). : × (5) In equation (6), represent Norm, This represents the transpose operation.

[0029] Step e, perform A along dimension m. After activation, the distribution weight of A is obtained. ; Step f, utilizing distribution weights right Perform weighted aggregation to obtain slot features. .

[0030] Step g: According to the number of slots s, sort the slot features Divided into e dedicated slot features ,in, Indicates the first Each unique slot feature; will Enter to the number Expert Module In this way, the first equation can be obtained using equation (7). Enhanced exclusive slot features : (6) In equation (6), It is the first j A linear mapping layer with doubled dimensions. express Activation function It is a linear mapping layer that compresses the j-th dimension.

[0031] Step h, along the number of slots s After splicing, the enhanced slot feature is obtained. .

[0032] Step i, along ( Dimension A After activation, the aggregate weight of A is obtained. ; and thus use right Weighted aggregation yields , Figure 4 The results show the t-SNE visualization of four experts using a hybrid expert layer with background probability modulation on three datasets: SIRSTv1, NUDT-SIRST, and IRSTD-1K. The results show that the hybrid expert layer with background probability modulation can effectively enhance the specialization of experts, thereby strengthening the model's learning ability.

[0033] Step 2.6.3, will After being processed in the decoding layer, the input undergoes further processing using the Sigmoid activation function to obtain... Predicted mask results .

[0034] Step 3, based on and The final loss of the infrared small target detection network is constructed using equation (1). : (1) In equation (1), For the segmentation loss function, The characteristic-level distillation loss function, As a response-level distillation loss function, this combination of feature-level distillation loss and response-level distillation loss can better facilitate the transfer of visual knowledge from the visible light domain to the infrared domain.

[0035] Step 3.1, construct the segmentation loss using equation (2). , This represents the binary cross-entropy loss function; (2) Step 3.2, construct the characteristic distillation loss using equation (3). : (3) In equation (3), Indicates the first The intermediate features of teachers output by the layer. Indicates the first Student intermediate features output by layer ; Represents the mean squared error loss function. Indicates the first A multi-level linear mapping layer that aligns the student intermediate features and teacher intermediate features output by dimension. This represents a token linear mapping layer that aligns the student global token with the teacher global token in terms of dimensions.

[0036] Step 3.3, construct the response-level distillation loss using equation (4). : (4) In equation (4), This is the distillation temperature coefficient. Let KL divergence loss function be used.

[0037] Step 4, Freeze and Then, the gradient descent method is used to backpropagate the infrared small target detection network, and the final loss is calculated. To optimize network parameters; obtain a trained infrared small target detection model; input the trained text. Input to The optimal text features are obtained through processing. And it is stored offline.

[0038] During the inference phase, the trained infrared visual encoder, the trained modulated infrared small target detector, and pre-stored... Small target detection is performed on the infrared image to be detected, and a segmentation threshold of 0.5 is used to obtain the optimal prediction mask result. An example of the partial detection result is shown below. Figure 5 As shown.

[0039] In this embodiment, an electronic device includes a memory and a processor. The memory stores a program that supports the processor in executing the above-described method, and the processor is configured to execute the program stored in the memory.

[0040] In this embodiment, a computer-readable storage medium stores a computer program, which is executed by a processor to perform the steps of the above method.

Claims

1. A method for detecting small infrared targets based on background probability modulation hybrid expert, characterized in that, Includes the following steps: Step 1: Acquire and preprocess infrared small target images with labeled information to obtain preprocessed infrared image samples. ,make The true value mask label is ,in, E, H, W These represent the number of channels, height, and width of the infrared image sample, respectively. Define a typical infrared background category set ;in, Indicates the first A typical infrared background category; K The number of typical infrared background categories; Step 2, construct an infrared small target detection network, including: CLIP text encoder CLIP image encoder Infrared visual encoder Modulated infrared small target detector, and for and Processing is performed to obtain Predicted mask results ;in, It uses the original text encoding structure that comes with the CLIP model. It adopts the standard ViT-B / 16 architecture. The ViT-Tiny / 16 structure is adopted. The modulated infrared small target detector consists of a coding layer, a background probability modulation hybrid expert layer, and a decoding layer. The coding layer consists of convolution and upsampling structures, and the decoding layer consists of convolution and upsampling structures. Step 3, based on and The final loss of constructing an infrared small target detection network ; Step 4, Freeze and Then, the gradient descent method is used to backpropagate the infrared small target detection network, and the final loss is calculated. The network parameters are optimized to obtain a trained infrared small target detection model; the trained text is then input. Input to The optimal text features are obtained through processing. And store offline; During the inference phase, the trained infrared visual encoder, the trained modulated infrared small target detector, and pre-stored... Small target detection is performed on the infrared image to be detected, and 0.5 is used as the segmentation threshold to obtain the optimal prediction mask result.

2. The infrared small target detection method based on background probability modulation hybrid expert as described in claim 1, characterized in that, Step 2 includes: Step 2.1, for Perform tokenization encoding to obtain the encoded category token. ,in, Indicates the number of category tokens. Indicates the dimension of the token; Concatenate the two tokens to be learned to obtain the text input. ,in, Represents a prefix token. Represents a suffix token. This represents channel concatenation, where n represents the number of prefix or suffix tokens. Step 2.2, will Input to CLIP text encoder After processing, text features are obtained. ; Step 2.3, will Input CLIP image encoder The process is performed to obtain the teacher intermediate features output from the 4th layer. The intermediate features of teachers output from layer 8 The intermediate features of teachers output from layer 12 And the teacher's global token extracted at the last layer. and will Input to linear mapping layer The image features of the teacher are then processed to obtain them. ; Step 2.4, Input infrared vision encoder The process is performed to obtain the student intermediate features output from the 4th layer. Student intermediate features output from layer 8 The intermediate features of students output from layer 12 And the student global token extracted at the last level. ; and will Input to linear mapping layer The student image features are then processed to obtain the student image features. ; Step 2.5, , respectively with Perform cosine similarity calculation to obtain the teacher's logical value. Logical value of students And use the Softmax activation function to Processing is performed to obtain the background probability. ; Step 2.6, The input is processed by the modulated infrared small target detector to obtain... Predicted mask results .

3. The infrared small target detection method based on background probability modulation hybrid expert as described in claim 2, characterized in that, Step 2.6 includes: Step 2.6.1, through the coding layer Feature extraction is performed to obtain deep visual tokens. , where m represents the number of deep visual tokens; Step 2.6.2, will and The inputs are fed into the background probability modulation mixing expert layer for processing to obtain the modulation token. ; Step 2.6.3, will After being processed in the decoding layer, the input undergoes further processing using the Sigmoid activation function to obtain... Predicted mask results .

4. The infrared small target detection method based on background probability modulation hybrid expert as described in claim 3, characterized in that, The background probability modulation hybrid expert layer in step 2.6.2 includes: Step a, first use the broadcast mechanism to... and The images are stitched together to obtain a visual representation with background probability awareness. ;right After normalization, the input is fed into the key mapping layer. The routing key matrix is ​​obtained through processing. ; Step b: Initialize a set of slot query matrices to be learned. Where e and s represent the number of experts and the number of slots for each expert, respectively; Step c, will Input to probability mapping layer The shape transformation is performed to obtain the background semantic query matrix. ; thereby and After concatenation, the final routing query matrix is ​​obtained. ; Step d, based on and The hyperspherical matched logic value is calculated using equation (5). : × (5) In equation (6), represent Norm, This represents the transpose operation; Step e, perform A along dimension m. After activation, the distribution weight of A is obtained. ; Step f, utilizing distribution weights right Perform weighted aggregation to obtain slot features. ; Step g: According to the number of slots s, sort the slot features Divided into e dedicated slot features ,in, Indicates the first Each unique slot feature; Will Enter to the number Expert Module In this way, the first equation can be obtained using equation (7). Enhanced exclusive slot features : (6) In equation (6), It is the first j A linear mapping layer with doubled dimensions. express Activation function It is a linear mapping layer that compresses the j-th dimension; Step h, along the number of slots s After splicing, the enhanced slot feature is obtained. ; Step i, along ( Dimension A After activation, the aggregate weight of A is obtained. Thus using right Weighted aggregation yields .

5. The infrared small target detection method based on background probability modulation hybrid expert as described in claim 2, characterized in that, Step 3 includes: Step 3.1, construct the segmentation loss using equation (1) : (1) In equation (1), This represents the binary cross-entropy loss function; Step 3.2, construct the characteristic distillation loss using equation (2). : (2) In equation (2), Indicates the first The intermediate features of teachers output by the layer. Indicates the first Student intermediate features output by layer ; Represents the mean squared error loss function. Indicates the first A multi-level linear mapping layer that aligns the student intermediate features and teacher intermediate features output by dimension. This represents a token linear mapping layer that aligns the student global token with the teacher global token in terms of dimensions. Step 3.3, construct the response-level distillation loss using equation (3). : (3) In equation (3), This is the distillation temperature coefficient. Let KL divergence loss function be used. Step 3.4, construct the final loss using equation (4). : (4)。 6. An electronic device, comprising a memory and a processor, characterized in that, The memory is used to store a program that supports a processor in executing the method of any one of claims 1-5, the processor being configured to execute the program stored in the memory.

7. A computer-readable storage medium storing a computer program thereon, characterized in that, The computer program is executed by the processor to perform the steps of the method according to any one of claims 1-5.

Images