A wheat diseased ear segmentation method based on perceptual enhanced gating network

Abstract

The present invention provides a method for segmenting diseased wheat ears based on a perception-enhanced gated network, comprising: obtaining a wheat image to be detected, wherein the wheat image to be detected includes a plurality of wheat ears; inputting the wheat image to be detected into a trained perception-enhanced gated network model to obtain a segmented image; the color of healthy wheat ears, the color of diseased wheat ears, and the color of a background area in the segmented image are different; the perception-enhanced gated network model integrates a dual-gating mechanism and a multi-scale dilated convolution block to achieve accurate segmentation of wheat ears. The method retains shallow local features in a semantic feature extraction branch through two different gating units while extracting deeper context features, obtaining a variety of feature details, so that the semantic segmentation model can still have excellent performance in complex scenes, thereby obtaining a very accurate segmented image.

Description

Technical Field

[0001] The present invention relates to the technical field of wheat ear disease detection, and in particular to a method for segmenting diseased wheat ears based on a perception-enhanced gated network. Background Art

[0002] As one of the most important crops in the world, wheat production plays a vital role in alleviating the global food crisis. However, wheat is susceptible to many biotic and abiotic stresses during the planting process, and these infections can seriously damage the ears, leaves, stems and other parts of the wheat. Among them, wheat fusarium wilt caused by Fusarium is one of the most difficult diseases in wheat production. It not only causes serious yield losses but also produces and accumulates toxins, which is extremely harmful to the health of humans and livestock. In the early stages of fusarium wilt infection, timely detection of the disease is conducive to taking control measures to minimize the spread of the disease, thereby increasing wheat yield and quality, increasing overall benefits, and ensuring global food security.

[0003] Early manual detection of wheat head rust was mainly carried out by experienced agricultural experts through measurement, sampling and analysis. This process relies on professional knowledge, experience and skills, so the reliability, objectivity and repeatability are limited. The current common field crop disease detection method usually uses the YOLO series of networks to locate the crop, and then separates the infected area through a simple threshold segmentation method. This method is redundant, complex and prone to errors. Summary of the invention

[0004] In view of the above defects in the prior art, the present invention provides a method for segmenting diseased wheat ears based on a perception-enhanced gated network to solve the technical problem in the prior art that diseased wheat ears are difficult to locate and segment.

[0005] To achieve the above-mentioned purpose and other related purposes, the present invention provides a method for segmenting diseased wheat ears based on a perceptual enhancement gated network, comprising: obtaining a wheat image to be detected, wherein the wheat image to be detected includes a plurality of wheat ears; inputting the wheat image to be detected into a trained perceptual enhancement gated network model to obtain a segmented image; the color of healthy wheat ears, the color of diseased wheat ears, and the color of the background area in the segmented image are different; the perceptual enhancement gated network model integrates a dual gating mechanism and a multi-scale dilated convolution block to achieve accurate segmentation of wheat ears.

[0006] In one embodiment of the present invention, the trained perceptual enhancement gated network model is obtained through the following steps: obtaining an original wheat image, wherein the original wheat image includes a plurality of wheat ears; manually labeling diseased wheat ears and healthy wheat ears in the original wheat image to obtain a mask image of the original wheat image; constructing a perceptual enhancement gated network model; using the original wheat image and its corresponding mask image as training samples, training the perceptual enhancement gated network model to obtain the trained perceptual enhancement gated network model.

[0007] In one embodiment of the present invention, the perception enhancement gated network model includes a starting unit, a feature extraction unit and an upsampling unit. The processing process of the perception enhancement gated network model can be expressed as: , where X l is the wheat image to be detected, X seg is the segmented image, H IU (), H FEU (), H UU () respectively represent the starting unit, the feature extraction unit and the upsampling unit.

[0008] In one embodiment of the present invention, the feature extraction unit includes a first branch and a second branch. The first branch extracts semantic features by cascading two different gating mechanisms, and the second branch improves segmentation accuracy by fusing feature information at multiple scales. The processing process of the feature extraction unit can be expressed as follows: , where X in1 and X out Respectively represent the input features and output features of the feature extraction unit, H a () and H b () represent the first branch and the second branch respectively.

[0009] In one embodiment of the present invention, the first branch is formed by stacking a plurality of DGB modules to gradually refine the extracted features. The processing process of the first branch can be expressed as: , where k = 1, 2, …, n, H k represents the kth DGB module, F k-1 and F k They represent the input features and output features of the kth DGB module, F0 is the output feature of the starting unit, and F n is the output feature of the first branch.

[0010] In one embodiment of the present invention, the DGB module includes an LGM submodule and an NGM submodule, and both the LGM submodule and the NGM submodule use a gating mechanism to improve the nonlinearity of the network so as to selectively focus on the required related features.

[0011] In one embodiment of the present invention, the processing of the LGM submodule and the NGM submodule can be respectively expressed as follows:

[0012] , ,

[0013] Where, X in2 represents the input features of the DGB module, X g1 , X g2 They represent the output features of the LGM submodule and the NGM submodule respectively, LN() represents layer normalization processing, Conv 1×1 () represents convolution processing, and LKC() represents convolution with a large-size kernel.

[0014] In one embodiment of the present invention, the second branch is obtained by improving the RFB-s module according to the following steps: replacing the 3×3 convolution in the RFB-s module with cascaded 1×3 convolution and 3×1 convolution.

[0015] To achieve the above-mentioned purpose and other related purposes, the present invention also provides an electronic device, including a processor, a memory and a communication bus; the communication bus is used to connect the processor and the memory; the processor is used to execute a computer program stored in the memory to implement a method provided in any one of the above embodiments.

[0016] To achieve the above-mentioned object and other related objects, the present invention further provides a computer-readable storage medium having a computer program stored thereon, wherein the computer program is used to enable a computer to execute the method provided in any one of the above-mentioned embodiments.

[0017] The beneficial effects of the present invention are as follows: the present invention proposes a method for segmenting diseased wheat ears based on a perception-enhanced gating network, which specifically extracts relatively rich feature information of wheat fusarium infection samples through a semantic feature extraction branch and a texture detail supplement branch of a parallel structure; at the same time, two different gating units retain shallow local features in the semantic feature extraction branch while extracting deeper contextual features, obtaining a variety of feature details, so that the semantic segmentation model can still have excellent performance in complex scenarios, thereby obtaining a very accurate segmented image. BRIEF DESCRIPTION OF THE DRAWINGS

[0018] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.

[0019] Figure 1 A flow chart of a segmentation method provided by an embodiment of the present invention;

[0020] Figure 2 A training flow chart of a perception-enhanced gated network model provided by an embodiment of the present invention;

[0021] Figure 3 A structural diagram of a perception enhancement gated network model provided by an embodiment of the present invention;

[0022] Figure 4 A structural diagram of a starting unit provided in one embodiment of the present invention;

[0023] Figure 5 A structural diagram of an up-sampling unit provided in one embodiment of the present invention;

[0024] Figure 6 A structural diagram of an LGM submodule provided in one embodiment of the present invention;

[0025] Figure 7 A structural diagram of an NGM submodule provided in one embodiment of the present invention;

[0026] Figure 8 A structural diagram of RFB-s' provided in one embodiment of the present invention;

[0027] Fig. 9 A comparison diagram of different segmentation methods provided in one embodiment of the present invention. DETAILED DESCRIPTION

[0028] The following describes the embodiments of the present invention through specific specific embodiments, and those skilled in the art can easily understand other advantages and effects of the present invention from the contents disclosed in this specification. It should be noted that the following embodiments and the features in the embodiments can be combined with each other without conflict. In addition to the specific methods, equipment, and materials used in the embodiments, according to the technical personnel in this technical field's grasp of the prior art and the records of the present invention, any methods, equipment, and materials of the prior art that are similar or equivalent to the methods, equipment, and materials described in the embodiments of the present invention can also be used to implement the present invention.

[0029] It should be understood that the terms used in the embodiments of the present invention are for describing specific embodiments rather than for limiting the scope of protection of the present invention. Unless otherwise defined, all technical and scientific terms used in the present invention have the same meaning as those generally understood by those skilled in the art.

[0030] In the following description, numerous details are discussed to provide a more thorough explanation of the embodiments of the present invention. However, it is obvious to those skilled in the art that the embodiments of the present invention can be implemented without these specific details. In some of the embodiments, well-known structures and devices are shown in the form of block diagrams rather than in detail to avoid making the embodiments of the present invention difficult to understand.

[0031] The flowchart and block diagram in the accompanying drawings illustrate the system architecture, functions and operations that may be implemented according to the methods and computer program products of various embodiments disclosed in the present invention. In this regard, each box in the flowchart or block diagram can represent a module, a program segment, or a part of a code, and the module, the program segment, or a part of the code contains one or more executable instructions for implementing the specified logical function. It should also be noted that in some alternative implementations, the functions marked in the box can also occur in a different order from the order marked in the accompanying drawings. For example, two boxes represented in succession can actually be executed substantially in parallel, and they can sometimes be executed in the opposite order, depending on the functions involved. It should also be noted that each box in the block diagram and / or the flowchart, and the combination of the boxes in the block diagram and / or the flowchart can be implemented with a dedicated hardware-based system that performs the specified function or operation, or can be implemented with a combination of dedicated hardware and computer instructions.

[0032] See also Figure 1 , Figure 1 A method for segmenting diseased wheat ears based on a perception-enhanced gated network is provided for one embodiment of the present invention, comprising: S100, obtaining a wheat image to be detected, wherein the wheat image to be detected includes a plurality of wheat ears; S200, inputting the wheat image to be detected into a trained perception-enrichment gated network model (i.e., Reception Enrichment Gate Network, referred to as REGN) to obtain a segmented image; the color of healthy wheat ears, the color of diseased wheat ears, and the color of the background area in the segmented image are different; the perception-enhanced gated network model integrates a dual-gating mechanism and a multi-scale dilated convolution block to achieve accurate segmentation of wheat ears.

[0033] This method uses a parallel structure of semantic feature extraction branches and texture detail supplement branches to specifically extract relatively rich feature information of wheat fusarium infection samples; at the same time, two different gating units retain shallow local features in the semantic feature extraction branch while extracting deeper contextual features, obtaining diverse feature details, so that the semantic segmentation model can still have excellent performance in complex scenarios, and thus obtain very accurate segmented images.

[0034] The size of the segmented image is the same as the size of the wheat image to be detected, so that the diseased wheat ears and healthy wheat ears in the segmented image correspond to the wheat ears in the wheat image to be detected, and further analysis can be performed later.

[0035] See also Figure 2 In a specific embodiment of the present invention, the trained perception enhancement gated network model is obtained through steps S201 to S204.

[0036] Step S201: Acquire an original wheat image, where the original wheat image includes a number of wheat ears.

[0037] Step S202, manually mark the diseased wheat ears and healthy wheat ears in the original wheat image to obtain the mask image of the original wheat image. The original wheat image and the mask image can be, for example, both RGB images, or other formats. In order to meet the resolution requirements of the network model, the resolutions of the original wheat image and the mask image can be set to a size of 512×512. The original wheat image can be marked to obtain the mask image, and then the size can be adjusted uniformly, or the size of the original wheat image can be adjusted first and then manually marked. When performing manual marking, for example, the ISAT_with_segment_anything annotation software can be selected for semi-automatic annotation, or other software can be used for auxiliary marking. During manual marking, both healthy wheat ears and diseased wheat ears are marked, and the remaining areas are defaulted to background areas.

[0038] Step S203: construct a perception enhancement gated network model. The specific structure of the network model will be described in detail later.

[0039] Step S204, using the original wheat image and its corresponding mask image as training samples, training the perception enhancement gated network model to obtain a trained perception enhancement gated network model. When constructing the sample, a variety of data enhancement techniques such as translation, rotation, and flipping can also be applied to enrich the training samples and improve the model performance. After obtaining the sample, it can be divided into a training set and a test set according to a certain ratio, and the ratio can be 4:1, for example. Then, the training set is used to train the perception enhancement gated network model. After the training is completed, the trained perception enhancement gated network model is tested using the test set. After the test is qualified, the trained perception enhancement gated network model is saved. If the test fails, the sample is re-acquired and then trained.

[0040] See also Figure 3 In a specific embodiment of the present invention, the perception enhancement gating network model includes an initial unit, a feature extraction unit, and an upsample unit. The processing process of the perception enhancement gating network model can be expressed as follows:

[0041] , where X l is the wheat image to be detected, X seg is the segmented image, H IU (), H FEU (), H UU () represent the starting unit, feature extraction unit and upsampling unit respectively.

[0042] See also Figure 4 In a specific embodiment of the present invention, the starting unit generally refers to the first processing unit of the network, which is responsible for receiving input data and performing preliminary feature extraction. In the perception enhancement gated network model of the present invention, the starting unit is responsible for receiving the input wheat image to be detected and performing preliminary feature extraction. In this embodiment, the starting unit is composed of three 3×3 convolution cascades.

[0043] See also Figure 5 In a specific embodiment of the present invention, the upsampling unit is used to ensure that the size of the output segmented image is the same as the size of the wheat image to be detected. When the wheat image to be detected is input into the trained perceptual enhancement gated network model for processing, the spatial resolution of the feature map will gradually decrease, and the upsampling unit is used to restore these low-resolution feature maps to a higher resolution. In this embodiment, the upsampling unit is composed of a 3×3 convolution and a transposed convolution cascade.

[0044] See also Figure 3In a specific embodiment of the present invention, the feature extraction unit includes a first branch and a second branch. The first branch extracts semantic features by cascading two different gating mechanisms, and the second branch improves segmentation accuracy by fusing feature information at multiple scales. The processing process of the feature extraction unit can be expressed as:

[0045] , where X in1 and X out Respectively represent the input features and output features of the feature extraction unit, H a () and H b () represent the first branch and the second branch respectively. The two branches are used to extract contextual semantic features and supplement texture details respectively. These branches use different gating mechanisms and the fusion of multi-scale information to enhance the feature expression ability of the network and improve the final segmentation accuracy.

[0046] See also Figure 3 In a specific embodiment of the present invention, the first branch is formed by stacking multiple DGB modules (Dual GateBlock) to gradually refine the extracted features. The processing process of the first branch can be expressed as: , where k = 1, 2, …, n, H k represents the kth DGB module, F k-1 and F k They represent the input features and output features of the kth DGB module respectively, F0 is the output feature of the starting unit, that is, the input feature X of the feature extraction unit in1 , F n It is the output feature of the first branch. The reason why multiple DGB modules are stacked to form the first branch is that this can gradually refine the extracted features and ensure the accuracy of the final segmentation.

[0047] See also Figure 6 and Figure 7 In a specific embodiment of the present invention, the DGB module includes an LGM submodule and an NGM submodule, and both the LGM submodule and the NGM submodule use a gating mechanism to improve the nonlinearity of the network so as to selectively focus on the required related features. Both the LGM submodule and the NGM submodule use a gating mechanism to improve the nonlinearity of the network, and can selectively focus on the required related features. These two complementary gating modules are cascaded to obtain sufficient feature extraction. This collaborative design enables the network to effectively learn the complementary information within the features, thereby solving the problem of limited information extraction ability of the existing segmentation network.

[0048] The processing of the LGM submodule and the NGM submodule can be expressed as follows:

[0049] ,

[0050] ,

[0051] Where, X in2 represents the input feature of the DGB module. For the first DGB module, it is the input feature X of the feature extraction unit. in1 , for the kth DGB module, it is the output of the k-1th DGB module. X g1 , X g2 They represent the output features of the LGM submodule and the NGM submodule respectively. LN() represents layer normalization processing. Conv 1×1 () represents convolution processing, and LKC() represents convolution with a large-size kernel.

[0052] The structure of LKC() is a prior art. Experiments have verified that the saturation point is reached when the kernel size increases to 7*7, and there will be no further performance gain when the kernel size continues to increase. The LKC used in this embodiment is a 7*7, R=2 dilated convolution, which can also be said to be a large kernel convolution with a size of 7*7.

[0053] A normalization layer is added before the LGM submodule and the NGM submodule to improve the generalization ability, and the modules are connected with residuals. The NGM submodule consists of the LGM submodule plus the LKC. The large-scale extended convolution enhances the receptive field of the model and extracts richer adjacent information.

[0054] See also Figure 8 In a specific embodiment of the present invention, the second branch is obtained by improving the RFB-s module according to the following steps: the 3×3 convolution in the RFB-s module is replaced by cascaded 1×3 convolution and 3×1 convolution. The improved RFB-s module is Figure 3 RFB-s' in.

[0055] RFBNet mainly wants to use some techniques to make a lightweight model a detector that achieves a good trade-off between speed and accuracy. It is inspired by the receptive field structure of human vision, Receptive Fields (RFs), and proposes a relatively novel RFblock (RFB) module to verify the relationship between the receptive field size and directionality in improving the discriminative robust features. RFBNet is built with SSD as the backbone network of VGG. It mainly adds dilated convolution on the basis of inception, thereby effectively increasing the receptive field. Overall, because it is improved based on the SSD network, the detection speed is still relatively fast, and there is a certain guarantee in terms of accuracy. The RFB module generally has two structures, RFB and RFB-s. In this embodiment, the second branch is improved by the RFB-s module, and its structure is as follows Figure 8 As shown, Figure 8 The dashed box in is the improved part. The second branch can further improve the segmentation accuracy by fusing feature information at multiple scales.

[0056] In order to illustrate the advantages of the segmentation method in the present invention compared with other segmentation methods in the prior art, further description is given below in combination with specific experimental data.

[0057] In the present invention, the construction, training and data processing of the model are carried out in a computer, the computer hardware environment is AMD EPYC 7642 48-core CPU and GTX3090 GPU, the software environment is the operating system of Ubuntu20.04, the compilation environment is PyCharm, and the deep learning framework is PyTorch; all subsequent training and testing are based on this platform. The present invention uses a drone to shoot images of wheat ears, and pre-processes them. The obtained images are used as data sets and divided into training sets and verification sets. The selected images are all RGB images of size 512×512.

[0058] In order to evaluate the segmentation performance of the segmentation method of the present invention, the present invention uses mean pixel accuracy (mPA), mean intersection over union (mIoU), mean Dice coefficient (mDice) and model size (Msize) as performance evaluation indicators of the present invention method.

[0059] mPA is the average pixel accuracy of the category, that is, first calculate the accuracy of the pixels that actually belong to category i in the predicted value of category i, then sum up all categories and take the average. Its expression is as follows:

[0060] ,

[0061] mIoU is the average intersection over union ratio, which is the ratio of the intersection and union of each type of predicted area result and the true value area result of the model, then sum and calculate the average. Its expression is as follows: ,

[0062] The mDice coefficient is a function that measures the similarity between two sets and is defined as twice the intersection divided by the sum of pixels. Its expression is as follows: ,

[0063] Msize represents the size of the optimal network model, which facilitates the direct screening of lightweight segmentation models.

[0064] In the above formulas, k represents the total number of model output classes, and k+1 includes empty classes or backgrounds; the confusion matrix is ​​shown in Table 1, pij represents the number of pixels belonging to the i-th class but misclassified as class j, p ji represents the number of pixels belonging to the jth class but misclassified as class i, p ii Represents the number of pixels correctly classified as infected with fusarium fusarium in wheat ears.

[0065] Table 1. Confusion matrix of classification results

[0066]

[0067] In order to ensure the consistency of the operating environment, 80% of the data from the wheat data set after wheat preprocessing is randomly selected as the network model training samples, and the rest are used as test samples. The batch size is set to 2, the maximum learning rate is set to 0.001, the minimum learning rate of the model defaults to 0.01 of the maximum learning rate, the learning rate decrease method is cosine annealing, and the weight is saved every 10 epochs, and a total of 150 epochs are trained.

[0068] In order to verify the ability of the perception-enhanced gated network model in the present invention to extract information about target features in the segmentation task, the scheme in the present invention is compared with some common semantic segmentation network methods in recent years (Deeplab v3+, HRNet, PSPNet, SegNet, SegFormer and FBSNet). The comparison results are shown in Table 2, where the best results are indicated in bold and the second best results are indicated by underscores.

[0069] Table 2 Comparison with common methods on wheat image segmentation tasks

[0070]

[0071] From the quantitative comparison results in Table 2, it can be seen that the present invention achieves the best semantic segmentation results in all cases because it uses a rich perception-enhanced semantic segmentation method to process the texture and detail information in the image.

[0072] The qualitative comparison results are shown in Fig. 9 ,It can be seen from the figure that no matter in the case of normal wheat ears or dense wheat ears, or in the case of shadow occlusion or exposure scene, the perception-enhanced gated network model proposed in the present invention can accurately segment the wheat ears.

[0073] An embodiment of the present invention further provides an electronic device, including a processor, a memory and a communication bus; the communication bus is used to connect the processor and the memory; the processor is used to execute a computer program stored in the memory to implement the above-mentioned segmentation method.

[0074] An embodiment of the present invention further provides a computer-readable storage medium on which a computer program is stored. The computer program is used to enable a computer to execute the above segmentation method.

[0075] The above embodiments are merely illustrative of the principles and effects of the present invention, and are not intended to limit the present invention. Anyone familiar with the art may modify or alter the above embodiments without departing from the spirit and scope of the present invention. Therefore, all equivalent modifications or alterations made by a person of ordinary skill in the art without departing from the spirit and technical concept disclosed by the present invention shall still be covered by the claims of the present invention.

Claims

1. A method for segmenting diseased wheat ears based on a perceptual enhanced gated network, characterized in that: include: Acquire a wheat image to be detected, wherein the wheat image to be detected includes a plurality of wheat ears; The wheat image to be detected is input into a trained perception enhancement gated network model to obtain a segmented image; the color of a healthy wheat ear, the color of a diseased wheat ear, and the color of a background area in the segmented image are different; the perception enhancement gated network model integrates a dual gating mechanism and a multi-scale dilated convolution block to achieve accurate segmentation of wheat ears; The perception enhancement gated network model includes a starting unit, a feature extraction unit, and an upsampling unit. The processing process of the perception enhancement gated network model can be expressed as: X seg =H UU (H FEU (H IU (X l ))), Where, X l is the wheat image to be detected, X seg is the segmented image, H IU (), H FEU (), H UU () respectively represent the starting unit, the feature extraction unit and the up-sampling unit; The feature extraction unit includes a first branch and a second branch. The first branch extracts semantic features by cascading two different gating mechanisms, and the second branch improves segmentation accuracy by fusing feature information at multiple scales. The processing process of the feature extraction unit can be expressed as follows: X out =H a (X in1 )+H b (X in1 ), Where, X in1 and X out Respectively represent the input features and output features of the feature extraction unit, H a () and H b () represent the first branch and the second branch respectively.

2. The method for segmenting diseased wheat ears based on a perception-enhanced gated network according to claim 1, characterized in that: The trained perception enhancement gated network model is obtained by the following steps: Acquire an original wheat image, wherein the original wheat image includes a plurality of wheat ears; Manually marking diseased wheat ears and healthy wheat ears in the original wheat image to obtain a mask image of the original wheat image; Construct a perception-enhanced gating network model; The original wheat image and its corresponding mask image are used as training samples to train the perception enhancement gated network model to obtain the trained perception enhancement gated network model.

3. The method for segmenting diseased wheat ears based on a perception-enhanced gated network according to claim 1, characterized in that: The first branch is formed by stacking multiple DGB modules to gradually refine the extracted features. The processing process of the first branch can be expressed as: F k =H k (F k-1 ), Where k = 1, 2, ..., n, H k represents the kth DGB module, F k-1 and F k They represent the input features and output features of the kth DGB module, F0 is the output feature of the starting unit, and F n is the output feature of the first branch.

4. The method for segmenting diseased wheat ears based on a perception-enhanced gated network according to claim 3, characterized in that: The DGB module includes an LGM submodule and an NGM submodule, and both the LGM submodule and the NGM submodule use a gating mechanism to improve the nonlinearity of the network so as to selectively focus on the required related features.

5. The method for segmenting diseased wheat ears based on perception-enhanced gated network according to claim 4, characterized in that: The processing procedures of the LGM submodule and the NGM submodule can be respectively expressed as follows: X g1 =Conv 1×1 (LN(X in2 ))·Conv 1×1 (LN(X in2 ))+X in2 , X g2 =LKC(Conv 1×1 (LN(X g1 )))·Conv 1×1 (LN(X g1 ))+X g1 , Where, X in2 represents the input features of the DGB module, X g1 , X g2 They represent the output features of the LGM submodule and the NGM submodule respectively, LN() represents layer normalization processing, Conv 1×1 () represents convolution processing, and LKC() represents convolution with a large-size kernel.

6. The method for segmenting diseased wheat ears based on a perception-enhanced gated network according to claim 1, characterized in that: The second branch is obtained by improving the RFB-s module according to the following steps: the 3×3 convolution in the RFB-s module is replaced by cascaded 1×3 convolution and 3×1 convolution.

7. An electronic device, characterized in that: The invention comprises a processor, a memory and a communication bus; the communication bus is used to connect the processor and the memory; the processor is used to execute a computer program stored in the memory to implement the method according to any one of claims 1 to 6.

8. A computer-readable storage medium, characterized in that: A computer program is stored thereon, and the computer program is used to enable a computer to execute the method as claimed in any one of claims 1 to 6.

Images