An insulator image quality evaluation method and system based on multi-task learning

Abstract

The present invention discloses an insulator image quality evaluation method and system based on multi-task learning, which relates to the technical field of insulator image evaluation. It includes: collecting high-definition insulator images, adding distortions with grade divisions, comprehensively using advanced objective evaluation algorithms for quality label annotation, constructing an insulator distortion image dataset, and through a multi-scale decoupled feature extraction method, fusing fine-grained features at different levels, encoding them through an encoder, inputting the obtained encoded features into the task interaction block of a task-aware Transformer decoder to capture the task interaction of each task, decoding the perceptual features of each task, and evaluating the insulator distortion images based on the MLP network. The present invention realizes the screening and preprocessing of distorted insulator components in the application of insulator fault diagnosis, and plays an important role in improving the accuracy of insulator fault diagnosis algorithms in transmission lines.

Description

Technical Field

[0001] The present invention relates to the technical field of insulator image quality evaluation, and in particular, to an insulator image quality evaluation method and system based on multi-task learning. Background Art

[0002] In a transmission line, there are numerous insulators and they are widely distributed. Due to being in the outdoor environment for a long time, insulators are easily affected by high voltage and complex climate conditions, and are prone to problems such as defects and cracks. High-efficiency and high-precision identification of insulator components in aerial images plays an important role in the timely diagnosis of transmission line faults. However, due to factors such as camera aging and damage, as well as external influences such as jitter during the shooting process and complex environments, the captured insulator images are distorted and degraded, seriously affecting the recognition accuracy of insulators and the defect detection accuracy. Therefore, there is an urgent need for a high-performance quality evaluation algorithm to monitor and evaluate the quality of the captured insulator images.

[0003] Subjective image quality evaluation is not only time-consuming and laborious, but also prone to causing visual fatigue to the human eye; objective image quality evaluation is a system that uses a computer to score a large number of distorted images, saving manpower and material resources and being able to quickly judge the images. The full-reference image quality evaluation algorithm requires the participation of a high-definition image. By comparing the statistical characteristics between the distorted image and the high-definition image, the gap or similarity between the two is calculated, and then a score is given. The semi-reference image quality evaluation scores the image by comparing some features of the high-definition image and the distorted image. However, in practical applications, it is often difficult to find the corresponding high-definition image. The no-reference image quality assessment (NR-IQA) algorithm that does not require a high-definition image is widely used in practice.

[0004] The insulator images collected by drones may be degraded due to bad weather, camera motion, etc., but in actual applications, there are no high-definition images for comparison, so only no-reference image quality evaluation algorithms can be used, which increases the difficulty of evaluation; the current public image quality evaluation algorithms have achieved good results in the evaluation of natural scene images, but once they are transferred to insulator images, they will not be consistent with human opinions, so it is necessary to develop an image quality evaluation algorithm for insulator images; there is currently no quality evaluation dataset for insulator images, and the creation of datasets requires a large number of images with human opinion ratings, and manual ratings consume a lot of manpower and time. At present, no-reference quality evaluation algorithms based on convolutional neural networks (CNNs) usually directly learn the mapping between images and subjective scores, but these algorithms tend to directly solve complex regression problems; and the global features used are not enough to capture complex distortions and cannot consider multi-scale distortion patterns. In order to solve the complex problem of end-to-end training, some studies divide image quality evaluation into several subtasks, and other subtasks are used to assist the image quality evaluation task. However, these multi-task learning methods mainly focus on shared features and lack the ability to interact between tasks.

[0005] It can be seen that there is a lack of reliable methods for quality evaluation of insulator images in the prior art. Therefore, better mining of the distortion characteristics of insulator images and realizing a reference-free image quality evaluation method that fully captures the distortion characteristics of insulator images are scientific problems to be solved by the present invention. In addition, how to evaluate the quality of insulator images and grade distortion without manual scoring and create a quality evaluation dataset based on insulator images is an urgent problem to be solved. Summary of the invention

[0006] In order to solve the above problems, the purpose of the present invention is to provide an insulator image quality assessment method and system based on multi-task learning, by constructing an insulator image quality assessment dataset, adding fine-grained feature fusion to enhance multi-scale distortion information, introducing a deformable mixer encoder module and a task-aware transformer decoder module, highlighting more information areas about different tasks, and improving the interactive ability of tasks, thereby completing the quality evaluation of insulator distorted images.

[0007] In order to achieve the above technical objectives, the present application provides an insulator image quality evaluation method based on multi-task learning, comprising the following steps:

[0008] Collect high-definition insulator images, add distortion with graded classification, and use objective evaluation algorithms to annotate quality labels to build an insulator distortion image dataset;

[0009] Based on the insulator distorted image dataset, through a multi-scale decoupled feature extraction method, after fusing fine-grained features at different levels, it is encoded by an encoder, and the obtained encoded features are input into the task interaction block of the task-aware transformer decoder;

[0010] Based on the task interaction block, capture the task interaction of each task, decode the perceptual features of each task, and evaluate the insulator distorted image according to the MLP network.

[0011] Preferably, during the process of quality label annotation, the MDSI algorithm is used to perform quality label annotation on the high-definition insulator images with added distortion.

[0012] Preferably, during the process of quality label annotation, for the insulator distorted images with quality label annotation, the Pearson linear correlation coefficient test is performed through the GMSD algorithm to eliminate incorrect labels.

[0013] Preferably, during the process of quality label annotation, according to the distortion level, the range of scalar quality scores is divided into discrete sub-intervals, and each sub-interval becomes a quality level as the quality grade label.

[0014] Preferably, before encoding by the encoder, the channels of each layer of feature maps of the Resnet50 network are divided into 4 groups. The features of the first group of channels are passed downwards, and the features of the second to fourth groups of channels are subjected to feature extraction through a 3×3 convolution, and after splicing and restoration, they are passed downwards. Then, a 1×1 convolution is used to fuse the information of the 4 groups of channels to achieve the extraction of multi-scale features.

[0015] Preferably, during the process of encoding by the encoder, the encoder is a deformable mixer encoder, including a linear layer for image feature dimensionality reduction, a channel-aware module for channel mixing through standard pointwise convolution, a first GELU activation + BatchNorm module, a spatial-aware module for obtaining the corresponding offsets of reference points through spatial context aggregation, a second GELU activation + BatchNorm module, a residual module, and a Reshape module for flattening image features. Among them, the residual module connects the outputs of the first GELU activation + BatchNorm module and the second GELU activation + BatchNorm module respectively.

[0016] Preferably, during the process of decoding by the task-aware transformer decoder, two encoded features are obtained through two deformable mixer encoders;

[0017] Two encoded features are input into the task-aware Transformer decoder for decoding. Among them, the task interaction block consists of a multi-head self-attention module (MHSA) and a small multi-layer perceptron (sMLP). The multi-head self-attention module (MHSA) is used to project the features and construct a self-attention strategy including query, key, and value matrices. The small multi-layer perceptron (sMLP) is used to generate interaction features for the quality scoring task and the quality rating task according to the self-attention strategy. The small multi-layer perceptron (sMLP) consists of a linear layer and a LayerNorm.

[0018] Preferably, in the process of obtaining the perceptual features, based on the interaction features, through two task query blocks, LayerNorm is applied in parallel to obtain the perceptual features of the decoding task.

[0019] Preferably, in the process of quality evaluation, based on the perceptual features, according to two independent MLP networks, by performing quality scoring prediction and distortion level prediction on the distorted insulator images, and according to the prediction results, the quality of the distorted insulator images is evaluated. Among them, the loss function uses a dynamic weight coefficient to determine the loss contribution of the two tasks, focuses on learning the distortion rating task at the initial stage of training, simulates the learning law from easy to difficult, takes the average of the image quality scores and distortion levels of each sub-block obtained to get the final quality score and distortion level of the entire image, and thus completes the quality evaluation of the distorted insulator images.

[0020] The present invention also provides an insulator image quality evaluation system based on multi-task learning, including:

[0021] A data acquisition module for acquiring high-definition insulator images;

[0022] A data processing module for adding distortions with grade divisions to the high-definition insulator images according to the objective evaluation algorithm to construct an insulator distorted image dataset, and fusing fine-grained features at different levels through a multi-scale decoupled feature extraction method, and then encoding through an encoder to obtain encoded features;

[0023] An evaluation module for inputting the encoded features into the task interaction block of the task-aware Transformer decoder, capturing the task interaction of each task, decoding the perceptual features of each task, and evaluating the distorted insulator images according to the MLP network.

[0024] The present invention discloses the following technical effects:

[0025] The present invention quickly constructs an insulator image quality evaluation dataset by getting rid of the subjective scoring method;

[0026] The present invention adopts a feature extraction method of multi-scale feature decoupling to fully aggregate distortion information at multiple scales; by adding a deformable mixer encoder module and a task-aware transformer decoder module to handle multi-task learning problems, it is convenient to fully mine the shared features and respective perceptual features between the distortion level prediction task and the quality score prediction task.

[0027] The present invention simultaneously performs quality assessment and distortion grading on the collected insulator images, so as to realize the screening and preprocessing of distorted insulator components in the application of insulator fault diagnosis, and plays an important role in improving the accuracy of the insulator fault diagnosis algorithm in the transmission line. Brief Description of the Drawings

[0028] In order to more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the following will briefly introduce the drawings required to be used in the embodiments. Obviously, the drawings in the following description are only some embodiments of the present invention. For those of ordinary skill in the art, without creative efforts, other drawings can also be obtained based on these drawings.

[0029] Figure 1 It is a flowchart of the insulator image quality evaluation method based on multi-task learning according to the present invention;

[0030] Figure 2 It is a network structure diagram of the insulator image quality evaluation method based on multi-task learning according to the present invention;

[0031] Figure 3 It is a schematic diagram of the multi-scale decoupling process according to the present invention;

[0032] Figure 4 It is a schematic diagram of the deformable mixer encoder according to the present invention;

[0033] Figure 5 It is a schematic diagram of the task interaction block according to the present invention;

[0034] Figure 6 It is a schematic diagram of the task query block according to the present invention. Detailed Embodiments

[0035] To make the objectives, technical solutions, and advantages of the embodiments of this application clearer, the technical solutions in the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings in the embodiments of this application. Obviously, the described embodiments are only some, but not all, of the embodiments of this application. The components of the embodiments of this application described and illustrated herein generally can be arranged and designed in a variety of different configurations. Therefore, the following detailed description of the embodiments of this application provided in the accompanying drawings is not intended to limit the scope of the claimed application, but merely represents selected embodiments of this application. All other embodiments obtained by those skilled in the art based on the embodiments of this application without creative efforts fall within the scope of protection of this application.

[0036] As Figure 1-6 shown, Embodiment 1: The present invention first constructs an insulator distorted image dataset and proposes an insulator image quality evaluation method based on multi-task learning. A quality grading prediction task is introduced to help optimize the regression task. Fine-grained feature extraction is performed using multi-scale decoupling, multi-scale distortion features are fused, and a deformable mixer encoder module and a task-aware transformer decoder module are introduced to enhance the multi-task learning ability, solving the problem in the current technology of lacking a reliable method for evaluating the quality of insulator images. The following is the specific implementation process of the present invention:

[0037] The insulator images collected by drones are degraded due to adverse weather conditions, camera movement, etc. There is an urgent need for a high-performance quality monitoring algorithm to monitor and evaluate the quality of the collected insulator images. The previously disclosed algorithms cannot consider multi-scale distortion patterns, and the captured global features are insufficient to extract complex distortions. To avoid directly solving complex regression problems, some scholars divide the quality evaluation task into multiple subtasks. However, these methods mainly focus on shared features and lack interaction between tasks. The present invention first constructs an insulator distorted image dataset, proposes to use multi-scale decoupling for fine-grained feature extraction and fusion, extracts the distortion information of each scale as accurately as possible, and introduces an effective deformable mixer encoder module and a task-aware transformer decoder to improve the interaction ability between the quality score prediction task and the distortion level prediction task. The flowchart of the insulator image quality evaluation method based on multi-task learning is as Figure 1 shown.

[0038] Firstly, high-definition insulator images and the objective full-reference image quality assessment algorithm - MDSI are used for fractional label annotation. To eliminate incorrect labels, Pearson linear correlation coefficient (PLCC) tests are performed on distorted images of the same content using another advanced full-reference image quality assessment algorithm - GMSD, constructing a large dataset of insulator distorted images that is independent of subjective scoring. Then, the insulator distorted images are divided into a training set and a test set, and the quality scores are classified into distortion levels. The distorted images in the training set and their corresponding scoring labels and distortion level labels are used to train the network model. After sufficient training, the distorted images in the test set are input into the model for testing to obtain the corresponding distortion levels and quality scores. Among them, the network structure diagram of the insulator image quality assessment method based on multi-task learning is as shown in Figure 2 shown.

[0039] The insulator image quality assessment method based on multi-task learning mainly consists of a multi-scale feature extraction network, an encoder, a decoder, and a multi-task prediction network.

[0040] (1) Multi-scale feature extraction network:

[0041] In the assessment of insulator image quality, different degrees of distortion need to be extracted, and there are often various features of different scales hidden in the distorted images. The residual image contains important information related to image quality. In view of this characteristic, the present invention proposes the multi-scale feature extraction network shown in Figure 2 wherein the network extracts four layers of distortion features from the Resnet50 basic network, downsamples them to the same size, and then splices and fuses them. Among them, the downsampling process is implemented by 3×3 convolution. In order to perform more fine-grained multi-scale feature extraction, the 3×3 convolution is decoupled into multiple scales, as shown in Figure 3 shown.

[0042] Firstly, the channels are grouped, and the number of groups scale is 4. The features of the first group are passed downwards, and the features of the second group are subjected to feature extraction through a 3×3 convolution. As a result, the receptive field of the feature extraction changes accordingly. By analogy, the receptive field of the subsequent groups becomes larger. Finally, the features of each group are spliced and restored, and 1×1 convolution is used again to fuse the channel information to achieve the extraction of multi-scale features in the same layer.

[0043] (2) Encoder:

[0044] To improve the feature extraction ability of the network, the present invention introduces two independent deformable mixer encoders to output deformable features for two prediction tasks.

[0045] The deformable mixer encoder is a hybrid of spatially-aware deformable spatial features and channel-aware positional features, adaptively providing a more effective receptive field and sampling spatial locations for each task. Taking the generation of deformed features for Task 1 as an example, as Figure 4 shown.

[0046] First, the channel dimension of the image feature X ∈ R H / 4×W / 4×C′ is reduced from C to a smaller dimension C' through a linear layer. The channel-aware module allows communication between different channels and is mixed through standard pointwise convolution (with a 1×1 convolution kernel); followed by GELU activation and BatchNorm; the spatial-aware module can model spatial context aggregation, and the image feature is fed into the convolution operator to learn the corresponding offsets Δ(i, j) for all reference points. This process can be written as:

[0047]

[0048] W2 is a deformable weight, and Δ (i，j) is the learnable offset.

[0049] Subsequently, GELU activation, BatchNorm, and residual connection are performed. The Reshape operation flattens the feature X ∈ R H / 4×W / 4×C ′ into a sequence R N×C′ (N = H / 4 × W / 4).

[0050] The output of the deformable mixer encoder is the specific features for two tasks, which can be used as the input for the subsequent decoder.

[0051] (3) Decoder:

[0052] To improve the ability of multi-task interaction, the present invention introduces a task-aware transformer decoder.

[0053] The task-aware transformer decoder includes a task interaction block and two task query blocks, which are used to capture task interaction features and make corresponding predictions respectively.

[0054] The task interaction block consists of a multi-head self-attention module (MHSA) and a small multi-layer perceptron (sMLP), and captures the task interaction of each task through the attention mechanism, as Figure 5 shown.

[0055] First, connect the two output deformed features from the deformable mixer encoder,

[0056]

[0057] where, X f ∈ R 2N×C′ is the fused feature.

[0058] To achieve efficient task interaction, the features are first projected into queries (Q), keys (K), and values (V) of dimension dk, and then a self-attention strategy is constructed:

[0059] Q = LN(X f ), X = LN(X f ), V = LN(X f )

[0060] X′ f = MHSA(Q, K, V)

[0061] where Q ∈ R N×C′ , K ∈ R N×C′ and V ∈ R N×C′ are the query, key, and value matrices respectively. The self-attention is calculated from Q, K, and V, and the specific calculation of the self-attention is as follows:

[0062]

[0063] Finally, the sMLP is used to generate the interaction features of the quality scoring task and the quality rating task, as shown in the following formula.

[0064]

[0065] where, is the task interaction feature, and the sMLP consists of a linear layer and a LayerNorm.

[0066] The task query block decodes the perceptual features of task 1 and task 2 from the predicted task interaction features respectively, as Figure 6 shown, taking the generation of the perceptual feature of task 1 as an example.

[0067] First, LayerNorm is applied in parallel to generate the query Q, key K, and value V. The deformed feature serves as the task query Q, and the task interaction feature serves as the key K and value V of the MHSA, and then a self-attention strategy is constructed:

[0068]

[0069]

[0070] The task perceptual feature is reshaped from R N×C′ (N = H / 4 × W / 4) through a reshaping operation and a residual connection is performed:

[0071]

[0072] (4) Quality fraction regression and distortion level classification network:

[0073] A quality rating prediction task simplified from a complex quality regression task is introduced to help optimize the regression task. Specifically, the range of scalar quality scores is divided into discrete sub-intervals, and each sub-interval becomes a quality level representing a specific quality level for the distortion rating prediction task. The network consists of two parts: an image quality score prediction module and an image distortion level prediction module, both implemented using simple multi-layer perceptron (MLP).

[0074] The steps of the insulator image quality evaluation method based on multi-task learning proposed in the present invention are as follows:

[0075] First step: Add five different types of distortions to the high-definition insulator images, each with five distortion levels. Use the objective evaluation algorithm to label the quality labels, and divide the range of scalar quality scores into discrete sub-intervals, so that each sub-interval becomes a quality level as the quality level label. Preprocess the distorted insulator images, and each distorted image is segmented into several small pieces of 224×224.

[0076] Second step: The feature extraction adopts a multi-scale decoupled feature extraction method. Finally, the fine-grained features at different levels are fused to form new features as the input of the deformable mixer encoder. Input the output result of the deformable mixer encoder into the task interaction block of the task-aware transformer decoder to capture the task interaction of each task; then input the interaction features into the task query block to decode the perceptual features of each task.

[0077] Third step: Use two independent MLP networks to predict the quality score and distortion level for each task obtained by the improved network. Among them, a dynamic weight coefficient is adopted in the loss function to determine the loss contribution of the two tasks, so that the model focuses on learning the distortion rating task at the initial stage of training, simulating the learning law from easy to difficult. Take the average of the image quality scores and distortion levels of each obtained block to get the final quality score and distortion level of the entire image, and then complete the quality evaluation of the distorted insulator image.

[0078] The present invention has the following features:

[0079] (1) The present invention first constructs a dataset of distorted insulator images, performs Gaussian filtering with different blur degrees, adds white noise with different levels, performs JPEG2000 compression and JPEG compression with different degrees, and adds motion blur with different blur degrees to high-definition insulator images to simulate various distortions in the acquisition process of insulator components in actual aerial images, forming a dataset for evaluating the quality of insulator images, and solving the problem of the lack of publicly available datasets for evaluating the quality of insulator images currently.

[0080] (2) The present invention abandons the subjective scoring method, uses a high-performance full-reference image quality assessment algorithm to score each distorted insulator image, and combines another high-performance full-reference image quality assessment algorithm to screen the labels and distorted images, realizing the rapid construction of labels in the dataset for evaluating the quality of insulator images.

[0081] (3) The present invention constructs an effective fine-grained feature extraction and fusion module, fully considers the influence of different-scale distortions, aggregates multi-scale distortion information, and solves the problem that most algorithms fail to fully consider multi-scale distortion patterns.

[0082] (4) The present invention also introduces a deformable mixer encoder module to capture more information regions related to each task; a task-aware transformer decoder is introduced to focus on the task awareness of each task, alleviating the problem of the lack of global modeling in CNNs, and improving the perception ability and interaction ability of the quality scoring task and the distortion classification task.

[0083] The present invention is described with reference to the flowcharts and / or block diagrams of methods, apparatuses (systems), and computer program products according to embodiments of the present invention. It should be understood that each process and / or block in the flowchart and / or block diagram, and the combination of processes and / or blocks in the flowchart and / or block diagram, can be implemented by computer program instructions. These computer program instructions can be provided to the processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing devices to generate a machine, such that the instructions executed by the processor of the computer or other programmable data processing devices generate means for implementing the functions specified in one Figure 1 one process or multiple processes and / or blocks Figure 1 one block or multiple blocks.

[0084] In the description of the present invention, it should be understood that the terms "first" and "second" are only used for descriptive purposes and cannot be construed as indicating or implying relative importance or implicitly specifying the quantity of the indicated technical features. Thus, the features defined with "first" and "second" may explicitly or implicitly include one or more of such features. In the description of the present invention, "a plurality of" means two or more unless otherwise specifically defined.

[0085] Obviously, those skilled in the art can make various changes and modifications to the present invention without departing from the spirit and scope of the present invention. Thus, if these modifications and variations of the present invention fall within the scope of the claims of the present invention and their equivalent technologies, the present invention is also intended to include these changes and modifications.

Claims

1. An insulator image quality evaluation method based on multi-task learning, characterized in that, It includes the following steps: Collect high-definition insulator images, add distortions with grade divisions, use objective evaluation algorithms for quality label annotation, and construct an insulator distortion image dataset; Based on the insulator distortion image dataset, through a multi-scale decoupled feature extraction method, after fusing fine-grained features at different levels, encode them through an encoder, and input the obtained encoded features into the task interaction block of the task-aware Transformer decoder; Based on the task interaction block, capture the task interaction of each task, decode the perceptual features of each task, and evaluate the insulator distortion image according to the MLP network; During the encoding process by the encoder, the encoder is a deformable mixer encoder, including a linear layer for image feature dimensionality reduction, a channel-aware module for channel mixing through standard pointwise convolutions, a first GELU activation + BatchNorm module, a spatial-aware module for obtaining the corresponding offsets of reference points through spatial context aggregation, a second GELU activation + BatchNorm module, a residual module, and a Reshape module for flattening image features. Among them, the residual module connects the outputs of the first GELU activation + BatchNorm module and the second GELU activation + BatchNorm module respectively; During the decoding process by the task-aware Transformer decoder, obtain two encoded features through two deformable mixer encoders; Input the two encoded features into the task-aware Transformer decoder for decoding. Among them, the task interaction block consists of a multi-head self-attention module MHSA and a small multi-layer perceptron sMLP. The multi-head self-attention module MHSA is used to project features and construct a self-attention strategy including query, key, and value matrices; the small multi-layer perceptron sMLP is used to generate interaction features for the quality scoring task and the distortion rating task according to the self-attention strategy. The small multi-layer perceptron sMLP consists of a linear layer and a LayerNorm layer; During the process of obtaining perceptual features, based on the interaction features, apply LayerNorm in parallel through two task query blocks to obtain the perceptual features of the decoding task; During the quality evaluation process, based on the perceptual features, according to two independent MLP networks, predict the quality score and distortion level of the insulator distortion image, and evaluate the quality of the insulator distortion image according to the prediction results. Among them, the loss function uses dynamic weight coefficients to determine the loss contribution of the two tasks. At the beginning of training, focus on learning the distortion rating task, simulate the learning law from easy to difficult, take the average of the image quality scores and distortion levels of each sub-block obtained to get the final quality score and distortion level of the entire image, and thus complete the quality evaluation of the insulator distortion image.

2. The method for evaluating the image quality of insulators based on multi-task learning according to claim 1, wherein: During the process of quality label annotation, the MDSI algorithm is used to perform quality label annotation on the high-definition insulator images with the added distortion.

3. The method for evaluating the image quality of insulators based on multi-task learning according to claim 2, wherein: During the process of quality label annotation, for the high-definition insulator images undergoing quality label annotation, Pearson linear correlation coefficient test is performed through the GMSD algorithm to eliminate incorrect labels.

4. The method for evaluating the image quality of insulators based on multi-task learning according to claim 3, wherein: During the process of quality label annotation, according to the distortion level, the range of scalar quality scores is divided into discrete sub-intervals, and each sub-interval becomes a quality level, serving as the quality grade label.

5. The method for evaluating the image quality of insulators based on multi-task learning according to claim 4, wherein: Before encoding through the encoder, the channels of the features of each layer of the Resnet50 network are divided into 4 groups. The features of the first group of channels are passed downwards. The features of the second to fourth groups of channels are subjected to feature extraction through a 3×3 convolution, and after splicing and restoration, they are passed downwards. Then, 1×1 convolution is used again to fuse the information of the 4 groups of channels for realizing the extraction of multi-scale features.

Images