A rice panicle infection segmentation method based on a receptive field progressive fusion network

Abstract

This invention provides a method for segmenting infected rice panicles based on a progressive fusion network of receptive fields. The method includes: acquiring an image of rice to be detected; processing the rice image using a trained rice false smut segmentation model to obtain a segmentation mask corresponding to the rice image; wherein the rice false smut segmentation model includes a feature extraction unit, a detail enhancement unit, and a segmentation head unit. In the feature extraction unit, a dynamic receptive field module processes some shallow feature maps to obtain enhanced shallow feature maps, thereby enhancing the perception ability of shallow features for small lesions and irregular edges. This method enhances shallow features through the dynamic receptive field module, improving the accuracy of capturing small lesions and irregular edges; the detail enhancement unit performs adaptive weighted fusion of multi-scale features, preserving edge details and incorporating deep semantics; the segmentation head unit restores spatial resolution, achieving clear segmentation of lesion boundaries and solving the problem of easily missed detection of small lesions.

Description

Technical Field

[0001] This invention relates to the field of crop disease detection and image processing technology, and in particular to a method for segmenting infected rice panicles based on a progressive fusion network of receptive fields. Background Technology

[0002] Rice is a vital staple crop globally, playing an irreplaceable role in ensuring food security and agricultural economic stability. Rice false smut, one of the major fungal diseases affecting rice in its later stages of growth, not only causes yield losses but also produces toxins that threaten food safety. Therefore, accurate and efficient detection in the early stages of the disease is of paramount importance.

[0003] With the development of UAV remote sensing technology and deep learning, disease monitoring methods based on semantic segmentation have become a research hotspot due to their wide coverage and high recognition accuracy. Existing segmentation networks typically employ an encoder-decoder structure, extracting deep semantic information through multiple downsampling operations. However, this process leads to the significant loss of shallow features containing details and location information, weakening the model's ability to perceive small lesions. In real, complex farmland scenarios, this can result in serious missed detections and incomplete edge segmentation. Furthermore, static convolutional kernels struggle to adapt to variations in lesion features under different textures and lighting conditions, further limiting the model's generalization ability. Summary of the Invention

[0004] In view of the above-mentioned deficiencies of the prior art, the present invention provides a method for segmenting infected rice panicles based on a progressive fusion network of the receptive field, so as to solve the technical problems of easy missed detection of small-sized lesions and incomplete edge segmentation in the prior art.

[0005] To achieve the above and other related objectives, this invention provides a method for segmenting infected rice panicles based on a progressive fusion network of receptive fields, comprising: acquiring an image of rice to be detected; processing the rice image using a trained rice false smut segmentation model to obtain a segmentation mask corresponding to the rice image; wherein, the rice false smut segmentation model comprises: a feature extraction unit, used to perform multi-level feature extraction on the rice image to obtain feature maps at different scales, and to process some shallow feature maps through a dynamic receptive field module to obtain enhanced shallow feature maps, the dynamic receptive field module adopting a multi-branch parallel dynamic dilated convolution structure, generating dynamic convolution kernels in real time according to input features, and performing dilated convolution calculations on different branches with different dilation rates to enhance the perception ability of shallow features for small lesions and irregular edges; a detail enhancement unit, used to receive the multi-scale feature maps output by the feature extraction unit, perform scale alignment and feature fusion on them to obtain fused multi-level features; and a segmentation head unit, used to upsample and integrate the fused multi-level features to gradually restore spatial resolution and generate the segmentation mask.

[0006] In one embodiment of the present invention, the feature extraction unit adopts a multi-layer coding structure, and performs downsampling processing on the rice image through multiple layers in sequence to extract feature maps of different scales layer by layer; wherein, at least one layer located in the shallow layer is provided with the dynamic receptive field module, which is used to enhance the receptive field of the feature map of the corresponding layer, while the other layers located in the deep layer are not provided with the dynamic receptive field module.

[0007] In one embodiment of the present invention, the feature extraction unit adopts a four-layer coding structure, wherein the first and second layers are respectively provided with the dynamic receptive field module, which is used to enhance the receptive field and adaptively reconstruct the shallow feature map of the corresponding layer; the third and fourth layers do not provide the dynamic receptive field module, and directly output the feature map of the corresponding layer.

[0008] In one embodiment of the present invention, the dynamic receptive field module adopts a multi-branch parallel structure, with each branch having a preprocessing submodule and a dynamic dilated convolution submodule. The preprocessing submodule is used to perform convolution operations on the input feature map to adjust the feature dimension or extract preliminary features. The dynamic dilated convolution submodule is used to generate dynamic convolution kernels in real time based on the input features and perform dilated convolution calculations. The output feature maps of each branch are fused and added to the input feature map through residual connections to obtain the output features of the dynamic receptive field module.

[0009] In one embodiment of the present invention, the dynamic receptive field module adopts a three-branch parallel structure, wherein: the preprocessing submodule of the first branch sequentially includes 1×1 convolution and 5×5 convolution, and the dilation rate of the dynamic dilated convolution submodule is set to 5; the preprocessing submodule of the second branch sequentially includes 1×1 convolution and 3×3 convolution, and the dilation rate of the dynamic dilated convolution submodule is set to 3; the preprocessing submodule of the third branch sequentially includes 1×1 convolution and 1×1 convolution, and the dilation rate of the dynamic dilated convolution submodule is set to 1.

[0010] In one embodiment of the present invention, the dynamic dilated convolution submodule performs the following operations: global average pooling is performed on the input feature map to extract the global context feature vector; the feature vector is input into an attention subnetwork composed of fully connected layers to generate weight coefficients; the weight coefficients are used to linearly weight and sum multiple preset base convolution kernels to synthesize a dynamic convolution kernel; and the synthesized dynamic convolution kernel is used to perform dilated convolution calculation with a preset dilation rate.

[0011] In one embodiment of the present invention, the detail enhancement unit includes a scale aggregation module, which is used to perform adaptive weighted fusion of the feature maps of the current level and its adjacent levels after scale alignment; wherein, the adjacent levels include the previous level and / or the next level, and the number of feature maps participating in the fusion is determined according to the actual existence of the levels.

[0012] In one embodiment of the present invention, the scale aggregation module performs the following operations: Each scale-aligned multi-scale feature map is passed through an independent 1×1 convolutional layer to map the number of channels of each feature map to 1, generating a corresponding single-channel feature scoring map; the single-channel feature scoring maps are concatenated along the channel dimension to construct a joint feature tensor; a Softmax activation function is applied to the joint feature tensor along the channel dimension to generate a normalized adaptive weight tensor; the adaptive weight tensor is separated along the channels to obtain spatial weight maps corresponding to each scale feature map; the aligned multi-scale feature maps are multiplied by their respective corresponding spatial weight maps, and the multiplication results are added element-wise to obtain a fused feature map.

[0013] In one embodiment of the present invention, the segmentation head unit receives the fused multi-level features, gradually restores the spatial resolution through multiple upsampling operations, and performs skip connections and fusion with the features of the corresponding level after each upsampling, and finally generates the segmentation mask through a convolutional layer.

[0014] The beneficial effects of this invention are as follows: This invention proposes a method for segmenting infected rice panicles based on a progressive fusion network of receptive fields. This method uses a dynamic receptive field module in the feature extraction unit and a multi-branch parallel dynamic dilated convolutional structure to generate dynamic convolutional kernels in real time according to the input features. This achieves receptive field enhancement and adaptive reconstruction of shallow features, significantly improving the accuracy of capturing small lesions and irregular edges. At the same time, the detail enhancement unit performs adaptive weighted fusion of multi-scale features, effectively solving the semantic conflicts between layers and integrating deep contextual information while preserving high-frequency edge details. The segmentation head unit gradually restores spatial resolution through progressive upsampling and skip connections, ultimately achieving clear segmentation of lesion boundaries and effective restoration of structural integrity. This solves the technical problems of easy missed detection of small lesions and incomplete edge segmentation in existing technologies. Attached Figure Description

[0015] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the accompanying drawings used in the description of the embodiments or the prior art will be briefly introduced below. The accompanying drawings are incorporated in and constitute a part of this specification, illustrating embodiments consistent with this application, and are used together with the description to explain the principles of this application. Obviously, the drawings described below are merely some embodiments of the present invention, and those skilled in the art can obtain other drawings based on these drawings without creative effort.

[0016] Figure 1 A flowchart of a segmentation method provided in an embodiment of the present invention;

[0017] Figure 2 This is a schematic diagram of a rice false smut segmentation model provided in an embodiment of the present invention;

[0018] Figure 3 This is an architectural diagram of a dynamic receptive field module provided in an embodiment of the present invention;

[0019] Figure 4 This is an architectural diagram of a dynamic dilated convolutional submodule provided in an embodiment of the present invention;

[0020] Figure 5 This is an architecture diagram of a scale aggregation module provided in an embodiment of the present invention;

[0021] Figure 6 This is a comparison diagram of different segmentation methods provided in an embodiment of the present invention. Detailed Implementation

[0022] The following specific embodiments illustrate the implementation of the present invention. Those skilled in the art can easily understand other advantages and effects of the present invention from the content disclosed in this specification. It should be noted that, unless otherwise specified, the following embodiments and features can be combined with each other. In addition to the specific methods, equipment, and materials used in the embodiments, based on the knowledge of the prior art and the description of the present invention by those skilled in the art, any prior art methods, equipment, and materials similar to or equivalent to the methods, equipment, and materials in the embodiments of the present invention can be used to implement the present invention.

[0023] It should be understood that the terminology used in the embodiments of this invention is for describing specific implementations and not for limiting the scope of protection of this invention. Unless otherwise defined, all technical and scientific terms used in this invention have the same meaning as commonly understood by one of ordinary skill in the art.

[0024] In the following description, numerous details are explored to provide a more thorough explanation of embodiments of the invention. However, it will be apparent to those skilled in the art that embodiments of the invention may be practiced without these specific details. In some embodiments, well-known structures and devices are shown in block diagram form rather than in detail to avoid obscuring embodiments of the invention.

[0025] The flowcharts and block diagrams in the accompanying drawings illustrate the architecture, functions, and operations that may be implemented in the methods and computer program products according to various embodiments of the present invention. In this regard, each block in a flowchart or block diagram may represent a module, segment, or portion of code containing one or more executable instructions for implementing the specified logical function. It should also be noted that in some alternative implementations, the functions indicated in the blocks may occur in a different order than those indicated in the drawings. For example, two consecutively indicated blocks may actually be executed substantially in parallel, and they may sometimes be executed in reverse order, depending on the functions involved. It should also be noted that each block in the block diagrams and / or flowcharts, and combinations of blocks in the block diagrams and / or flowcharts, can be implemented using a dedicated hardware-based system that performs the specified function or operation, or using a combination of dedicated hardware and computer instructions.

[0026] Please see Figure 1 , Figure 1 An embodiment of the present invention provides a method for segmenting infected rice panicles based on a progressive fusion network of receptive fields, comprising: (1) acquiring an image of rice to be detected; and (2) processing the rice image using a trained rice false smut segmentation model to obtain a segmentation mask corresponding to the rice image. The rice false smut segmentation model is as follows: Figure 2As shown, it includes a Feature Extraction Unit (FEU), a Detail Enhancement Unit (DEU), and a Segmentation Head Unit (SHU). The FEU performs multi-level feature extraction on the rice image, obtaining feature maps at different scales. A Dynamic Receptive Field Block (DRFB) is used to process some shallow feature maps to obtain enhanced shallow feature maps. The DRFB employs a multi-branch parallel dynamic dilated convolution structure, generating dynamic convolution kernels in real-time based on the input features. Different dilation rates are used for dilated convolution calculations on different branches to enhance the shallow features' ability to perceive small lesions and irregular edges. The Detail Enhancement Unit receives the multi-scale feature maps output by the FEU, performs scale alignment and feature fusion, obtaining fused multi-level features. The SHU upsamples and integrates the fused multi-level features, gradually restoring spatial resolution and generating a segmentation mask. The processing of this model can be represented as follows: X in and X out H corresponds to the input and output features of the model, respectively. FE H DE and H SH These correspond to the feature extraction unit, detail enhancement unit, and segmentation head unit, respectively.

[0027] In this embodiment, the rice blast disease segmentation model adopts an end-to-end training method, with an input of a 512×512 pixel RGB image and an output of a pixel-level segmentation mask of the same resolution. The feature extraction unit gradually extracts multi-scale features through a multi-layer coding structure. The shallow features are enhanced by a dynamic receptive field module, preserving rich edge and texture details. The detail enhancement unit performs adaptive weighted fusion of features at each level through a scale aggregation module, effectively fusing shallow spatial information and deep semantic information. The segmentation head unit gradually restores resolution through multiple upsampling and introduces skip connections to compensate for the details lost during upsampling. This structural design enables the model to accurately identify tiny lesions even in complex paddy field backgrounds, significantly improving the accuracy and robustness of segmentation.

[0028] Understandably, the rice false smut segmentation model is based on... Figure 2After the model is built, it needs to be trained. The training process may include the following steps: Large-size rice images are collected using a drone, cropped to 512×512 pixels, and then, under expert guidance, diseased and healthy areas are finely labeled to generate pixel-level mask labels. These labels are then divided into training and validation sets at a 4:1 ratio. During training, data augmentation strategies such as random flipping, rotation, and color jitter are used to improve the model's generalization ability. The model is jointly optimized using the cross-entropy loss function and the Dice loss function, with parameters updated using the Adam optimizer. The initial learning rate is set to 0.001, the training epochs are 200, and the batch size is set to 8. The model with the best performance on the validation set is saved as the final segmentation model during training.

[0029] Understandably, the original drone-captured images of rice paddies are too large to be directly input into the segmentation network for processing. Therefore, the large images need to be cropped into fixed-size sub-images of 512×512 pixels (or other sizes) to meet the network's input requirements. The cropped images retain the local details of the original images, facilitating the model's learning of the texture and edge features of lesions. The same cropping and preprocessing strategy is used for the images to be detected to ensure consistency with the training data distribution.

[0030] In a specific embodiment of the present invention, the feature extraction unit adopts a multi-layer coding structure, performing downsampling processing on the rice image layer by layer to extract feature maps of different scales. At least one layer in the shallow layers is equipped with a dynamic receptive field module to enhance the receptive field of the feature map at that layer, while the remaining layers in the deeper layers do not have a dynamic receptive field module. In this embodiment, the feature extraction unit adopts a multi-layer coding structure, expanding the receptive field and extracting deep semantic features through layer-by-layer downsampling. Shallow feature maps contain rich spatial detail information; setting a dynamic receptive field module to enhance them can effectively improve the perception of small lesions. Deep feature maps are mainly responsible for semantic abstraction; this module is not set in deep feature maps to reduce computational overhead and maintain feature purity.

[0031] In a specific embodiment of the present invention, the feature extraction unit adopts a four-layer coding structure. The first and second layers each have dynamic receptive field modules for enhancing the receptive field and adaptively reconstructing the shallow feature maps of the corresponding layers. The third and fourth layers do not have dynamic receptive field modules and directly output the feature maps of the corresponding layers. In this embodiment, the feature extraction unit adopts a four-layer coding structure. The first and second layers are shallow layers, outputting feature maps with high resolution and containing rich edge and texture information. Dynamic receptive field modules are set up for receptive field enhancement and adaptive reconstruction, enabling shallow features to more accurately capture minute lesions and irregular edges. The third and fourth layers are deep layers, outputting feature maps with lower resolution, primarily responsible for semantic information extraction, directly outputting the features of the corresponding layers to avoid redundant computation.

[0032] Please see Figure 3 In a specific embodiment of the present invention, the dynamic receptive field module adopts a multi-branch parallel structure. Each branch is equipped with a preprocessing submodule and a dynamic dilated convolution submodule. The preprocessing submodule performs convolution operations on the input feature map to adjust the feature dimensions or extract preliminary features. The dynamic dilated convolution submodule generates dynamic convolution kernels in real time based on the input features and performs dilated convolution calculations. The output feature maps of each branch are fused and added to the input feature map through residual connections to obtain the output features of the dynamic receptive field module. In this embodiment, the dynamic receptive field module adopts a multi-branch parallel structure, capturing multi-scale contextual information through different branches. Each branch first adjusts the feature dimensions and extracts preliminary features through the preprocessing submodule, and then the dynamic dilated convolution submodule generates dynamic convolution kernels in real time based on the input features and performs dilated convolution calculations, enabling the convolution kernels to adaptively adapt to different lesion features. After the multi-branch outputs are fused, they are added to the original input through residual connections, which enhances the feature representation capability and maintains the stability of gradient propagation.

[0033] See also Figure 3 In a specific embodiment of the present invention, the dynamic receptive field module adopts a three-branch parallel structure, wherein: the preprocessing submodule of the first branch includes 1×1 convolution and 5×5 convolution in sequence, and the dilation rate of the dynamic dilated convolution submodule is set to 5; the preprocessing submodule of the second branch includes 1×1 convolution and 3×3 convolution in sequence, and the dilation rate of the dynamic dilated convolution submodule is set to 3; the preprocessing submodule of the third branch includes 1×1 convolution and 1×1 convolution in sequence, and the dilation rate of the dynamic dilated convolution submodule is set to 1.

[0034] In this embodiment, the dynamic receptive field module adopts a three-branch parallel structure. Each branch uses different combinations of preprocessed convolutional kernels and dilation rates to achieve differentiated extraction of receptive fields at multiple scales. The first branch uses a 5×5 convolution with a dilation rate of 5 to capture a large range of contextual information; the second branch uses a 3×3 convolution with a dilation rate of 3 to focus on medium-scale features; and the third branch uses a 1×1 convolution with a dilation rate of 1 to preserve local details. The three branches work together to enable the model to simultaneously perceive lesion features at different scales, significantly improving the ability to identify small-sized and irregularly shaped lesions.

[0035] Please see Figure 4In a specific embodiment of the present invention, the dynamic dilated convolution submodule performs the following operations: (1) global average pooling (GAP) on the input feature map to extract the global context feature vector; (2) inputting the feature vector into an attention subnetwork (FC+ReLU+FC+Softmax) composed of fully connected layers, and mapping to generate weight coefficients; (3) using the weight coefficients to perform linear weighted summation on a preset set of base convolution kernels to synthesize a dynamic convolution kernel; (4) using the synthesized dynamic convolution kernel to perform dilated convolution calculation with a preset dilation rate. The core of the dynamic dilated convolution submodule lies in dynamically generating convolution kernels based on input features. First, global context information of the input features is extracted through global average pooling, and then weight coefficients are generated through mapping by the attention subnetwork. The weight coefficients are then used to perform linear weighted summation on a preset set of base convolution kernels to synthesize a dynamic convolution kernel specific to the current input features. This mechanism gives the network the ability to dynamically extract features for "a thousand images, a thousand faces", enabling the model to adaptively process lesion images under different textures and lighting conditions, effectively overcoming the problem of poor feature adaptability of static convolution kernels.

[0036] Specifically, the processing procedure of the dynamic receptive field module described above can be expressed by the following formula:

[0037] ,

[0038] ,

[0039] ,

[0040] ,

[0041] Where B1 is the first branch, B2 is the second branch, and B3 is the third branch. Represents a 1×1 convolution. Represents a 3×3 convolution. Represents a 5×5 convolution; Representative expansion rate The dynamic dilated convolution operation is defined by the following formula:

[0042] ,

[0043] In the formula, X represents the input feature of the dynamically dilated convolutional submodule, # r This represents a convolution operation with an expansion rate of r. W is the k-th dynamic weight coefficient generated by the attention subnet from the input feature X. k Let K be the k-th preset base convolution kernel, where K is the total number of base convolution kernels.

[0044] In a specific embodiment of the present invention, the detail enhancement unit includes a scale aggregation block (SAB). The SAB is used to adaptively weightedly fuse the feature maps of the current level and its adjacent levels after scale alignment. Adjacent levels include the previous level and / or the next level, and the number of feature maps participating in the fusion is determined based on the actual existence of the levels. The SAB is used to fuse the feature maps of adjacent levels, achieving complementarity between spatial details and semantic information. For the current level, the module receives the downsampled and aligned features of the previous level, the original features of the current level, and the upsampled and aligned features of the next level, integrating the three information through adaptive weighted fusion. For top or bottom boundary levels, features of non-existent adjacent levels are automatically omitted, ensuring the module's universality across different levels. This design effectively solves the semantic conflict problem during multi-level feature fusion.

[0045] Please see Figure 5 In a specific embodiment of the present invention, the scale aggregation module performs the following operations: (1) passing the scale-aligned multi-scale feature maps through independent 1×1 convolutional layers to map the number of channels of each feature map to 1, generating corresponding single-channel feature scoring maps; (2) splicing the single-channel feature scoring maps along the channel dimension to construct a joint feature tensor; (3) applying the Softmax activation function to the joint feature tensor along the channel dimension to generate a normalized adaptive weight tensor; (4) separating the adaptive weight tensor along the channel to obtain the spatial weight map corresponding to each scale feature map; (5) multiplying the aligned multi-scale feature maps with their respective corresponding spatial weight maps, and adding the multiplication results element by element to obtain the fused feature map.

[0046] In this embodiment, the scale aggregation module first generates single-channel score maps for feature maps at each scale using independent 1×1 convolutional layers. These score maps reflect the dependence of different spatial locations on features at each level. Then, the score maps are concatenated and normalized using Softmax to generate spatial weight tensors, ensuring that the sum of weights at each location is 1, achieving pixel-level dynamic weight allocation. Finally, the aligned feature maps are multiplied by their corresponding weights and added element-wise to complete adaptive weighted fusion. This mechanism enables the model to effectively incorporate deep contextual information while preserving high-frequency edge details, significantly improving the ability to clearly segment lesion boundaries and restore structural integrity.

[0047] by Figure 2 Taking a scale aggregation module with three inputs as an example, it receives a feature map from the feature extraction unit that has been upsampled or downsampled and scale-aligned, denoted as... F l l and ,in This represents the feature map after downsampling and alignment with the target scale. This represents the feature map after horizontal concatenation, where the original scale is the same as the target scale. This represents the feature map after upsampling and alignment with the target scale.

[0048] The specific fusion process of the scale aggregation module is as follows: (1) Multi-scale feature scoring generation: The feature map after scale alignment , and (C channels) are input into three independent 1×1 convolutional layers. The convolutional layer maps the number of channels of each feature map from C to 1, thus directly obtaining three independent single-channel feature scoring maps. (2) Scoring map joint: The three single-channel feature scoring maps are concatenated along the channel dimension to construct a joint feature tensor containing scoring information of all levels, with a channel number of 3. (3) Spatial weight normalization: The Softmax activation function is applied to the joint feature tensor along the channel dimension to generate a normalized adaptive weight tensor. (4) Weight parsing: The normalized weight tensor is separated along the channel to obtain the final three spatial weight maps. (5) Weighted fusion output: Multiply the aligned feature maps by their respective weight maps, and then add the results element by element to obtain the final fusion output. The resolution of this output is consistent with the "target resolution" selected in "Scale Alignment", i.e., the resolution of level l. This fusion output operation can be expressed as follows: at any spatial location (i,j), the output features of SAB after fusion at level l can be obtained. :

[0049] .

[0050] Please see Figure 2 In one specific embodiment of the present invention, the segmentation head unit receives the fused multi-level features, gradually restores the spatial resolution through multiple upsampling operations, and after each upsampling, fuses them with the features of the corresponding level through skip connections, finally generating a segmentation mask through a convolutional layer. The segmentation head unit receives the multi-level fused features output by the detail enhancement unit, gradually restores the spatial resolution through multiple upsampling operations. After each upsampling, it fuses them with the features of the corresponding level in the feature extraction unit through skip connections, supplementing the spatial detail information lost during the upsampling process, and finally generating a segmentation mask with the same resolution as the input image through a convolutional layer, achieving end-to-end pixel-level lesion segmentation.

[0051] It should be noted that the steps of the various methods described above are only for clarity. In practice, they can be combined into one step or some steps can be split into multiple steps. As long as they contain the same logical relationship, they are all within the scope of protection of this application. Adding insignificant modifications or introducing insignificant designs to the algorithm or process, but without changing the core design of the algorithm and process, are also within the scope of protection of this patent.

[0052] To verify the advantages of the segmentation method proposed in this invention compared with existing technologies, the following comparative analysis is conducted based on specific experimental data.

[0053] All experiments were conducted on a unified computer platform, including model building, training, and data processing. The hardware environment consisted of an AMD EPYC 7V12 32-core processor and an RTX 4080 GPU; the software environment was an Ubuntu 22.04 LTS operating system; programming was based on the PyCharm integrated development environment; and the deep learning framework was PyTorch 1.12. All training and testing processes were performed within this environment. This experiment used rice panicle images collected by drones as the dataset. After preprocessing, these images were uniformly cropped to 512×512 pixel RGB images and randomly divided into training and validation sets at a 4:1 ratio.

[0054] To evaluate the segmentation performance of the segmentation method of the present invention, the present invention uses the mean intersection-over-union (mIoU) and mean pixel accuracy (mPA) as performance evaluation indicators of the method.

[0055] mPA is the class average pixel accuracy, which is calculated by first determining the accuracy of pixels that truly belong to class i among the predicted values ​​for class i, then summing the values ​​for all classes and taking the average. Its expression is as follows:

[0056] ,

[0057] mIoU is the average intersection-union ratio, which is calculated by summing the ratios of the intersection and union of the predicted and actual regional results for each type of region, and then averaging them. Its expression is as follows:

[0058] ,

[0059] In the above formulas, k represents the total number of output classes of the model, and k+1 includes the empty class or background; the confusion matrix is ​​shown in Table 1, p ij p represents the number of pixels that belong to class i but are misclassified as class j. ji p represents the number of pixels that belong to class j but are misclassified as class i. iiThis represents the number of pixels that are correctly classified as rice panicle blast infection.

[0060] Table 1: Confusion matrix of classification results.

[0061]

[0062] To verify the effectiveness of the progressive fusion network model of receptive field in this invention for rice blast segmentation, a comparative experiment was conducted with some commonly used semantic segmentation networks in recent years (UNet, SegNet, DeepLabv3+, HRNet, Segformer, FBSNet). The results are shown in Table 2.

[0063] Table 2: Performance comparison of different models in semantic segmentation tasks.

[0064]

[0065] The quantitative comparison results in Table 2 show that the segmentation model provided by this invention exhibits excellent performance in the lesion and ear segmentation task. It significantly enhances the ability to identify small lesion targets through the receptive field block, and the progressive fusion of multi-scale features effectively improves the ability to restore details. As a result, the overall model achieves the best performance in both key evaluation indicators, with mIoU reaching 80.41% and mPA reaching 89.78%, which is significantly better than other comparative models.

[0066] like Figure 6 The qualitative comparison results show that the segmentation model proposed in this invention can still achieve accurate edge recognition and region segmentation when faced with tiny lesions whose color is very close to the background. The model effectively enhances the visibility of lesion targets in low-contrast areas and significantly improves the response intensity and continuity of small target boundaries. In complex backgrounds, the model can not only capture extremely fine lesion contours but also suppress background noise interference, avoiding edge breakage or blurring. It exhibits superior performance, especially in preserving edge details of tiny lesions, enhancing weak edges, and connecting fragmented areas, thus achieving high-integrity reconstruction and high-precision segmentation of fine lesion structures.

[0067] In summary, this invention effectively addresses the issues of feature loss and missed detection in small-sized lesion detection in existing semantic segmentation networks through the collaborative design of a dynamic receptive field module and a scale aggregation module. The dynamic receptive field module endows the network with dynamic feature extraction capabilities, enhancing adaptive perception of lesions under different textures and lighting conditions; the scale aggregation module achieves efficient fusion of multi-level features, incorporating deep semantic information while preserving edge details. Experimental results show that this invention significantly outperforms existing mainstream segmentation methods in both mIoU and mPA, providing reliable technical support for the accurate monitoring of rice false smut.

[0068] The above embodiments are merely illustrative of the principles and effects of the present invention and are not intended to limit the invention. Any person skilled in the art can 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 those skilled in the art without departing from the spirit and technical concept disclosed in the present invention should still be covered by the claims of the present invention.

Claims

1. A method for segmenting infected rice panicles based on a progressive fusion network of receptive fields, characterized in that, include: Acquire images of the rice plants to be inspected; The rice image is processed using a trained rice false smut segmentation model to obtain the segmentation mask corresponding to the rice image; The rice false smut segmentation model includes: The feature extraction unit is used to perform multi-level feature extraction on the rice image to obtain feature maps at different scales. The dynamic receptive field module processes some shallow feature maps to obtain enhanced shallow feature maps. The dynamic receptive field module adopts a multi-branch parallel dynamic dilated convolution structure, generates dynamic convolution kernels in real time according to the input features, and performs dilated convolution calculations on different branches with different dilation rates to enhance the ability of shallow features to perceive small lesions and irregular edges. The detail enhancement unit is used to receive the multi-scale feature map output by the feature extraction unit, perform scale alignment and feature fusion on it, and obtain fused multi-level features; and The segmentation head unit is used to upsample and integrate the fused multi-level features to gradually restore the spatial resolution and generate the segmentation mask.

2. The rice infection panicle segmentation method based on a progressive fusion network of receptive fields according to claim 1, characterized in that, The feature extraction unit adopts a multi-layer coding structure, and performs downsampling processing on the rice image through multiple layers in sequence to extract feature maps of different scales layer by layer; wherein, at least one layer in the shallow layer is provided with the dynamic receptive field module, which is used to enhance the receptive field of the feature map of the corresponding layer, while the other layers in the deep layer are not provided with the dynamic receptive field module.

3. The rice infection panicle segmentation method based on a progressive fusion network of receptive fields according to claim 2, characterized in that, The feature extraction unit adopts a four-layer coding structure, wherein the first and second layers are respectively equipped with the dynamic receptive field module, which is used to enhance the receptive field and adaptively reconstruct the shallow feature map of the corresponding layer; the third and fourth layers do not have the dynamic receptive field module and directly output the feature map of the corresponding layer.

4. The rice infection panicle segmentation method based on a progressive fusion network of receptive fields according to claim 1, characterized in that, The dynamic receptive field module adopts a multi-branch parallel structure, with each branch having a preprocessing submodule and a dynamic dilated convolution submodule, wherein: The preprocessing submodule is used to perform convolution operations on the input feature map to adjust the feature dimensions or extract preliminary features; The dynamic dilated convolution submodule is used to generate dynamic convolution kernels in real time based on input features and perform dilated convolution calculations. After the output feature maps of each branch are fused, they are added to the input feature map through residual connection to obtain the output features of the dynamic receptive field module.

5. The rice infection panicle segmentation method based on a progressive fusion network of receptive fields according to claim 4, characterized in that, The dynamic receptive field module adopts a three-branch parallel structure, wherein: The preprocessing submodule of the first branch includes 1×1 convolution and 5×5 convolution in sequence, and the dilation rate of the dynamic dilated convolution submodule is set to 5. The preprocessing submodule of the second branch includes 1×1 convolution and 3×3 convolution in sequence, and the dilation rate of the dynamic dilated convolution submodule is set to 3. The preprocessing submodule of the third branch includes 1×1 convolution and 1×1 convolution in sequence, and the dilation rate of the dynamic dilated convolution submodule is set to 1.

6. The rice infection panicle segmentation method based on a progressive fusion network of receptive fields according to claim 4 or 5, characterized in that, The dynamic dilated convolution submodule performs the following operations: Global average pooling is performed on the input feature map to extract the global context feature vector; The feature vector is input into an attention subnetwork composed of fully connected layers, and weight coefficients are generated by mapping. The weight coefficients are used to perform linear weighted summation on a preset set of base convolutional kernels to synthesize a dynamic convolutional kernel; Dilated convolutions are performed using synthesized dynamic convolution kernels with a preset dilation rate.

7. The rice infection panicle segmentation method based on a progressive fusion network of receptive fields according to claim 1, characterized in that, The detail enhancement unit includes a scale aggregation module, which is used to adaptively weight and fuse the feature maps of the current level and its adjacent levels after scale alignment; wherein, the adjacent levels include the previous level and / or the next level, and the number of feature maps participating in the fusion is determined according to the actual existence of the levels.

8. The rice infection panicle segmentation method based on a progressive fusion network of receptive fields according to claim 7, characterized in that, The scale aggregation module performs the following operations: After scale alignment, the multi-scale feature maps are passed through independent 1×1 convolutional layers to map the number of channels of each feature map to 1, thereby generating the corresponding single-channel feature scoring map. The feature scoring maps of each single channel are concatenated along the channel dimension to construct a joint feature tensor; Apply the Softmax activation function to the joint feature tensor in the channel dimension to generate a normalized adaptive weight tensor; The adaptive weight tensor is separated along the channel to obtain the spatial weight map corresponding to the feature map at each scale. The aligned multi-scale feature maps are multiplied by their respective spatial weight maps, and the multiplication results are added element by element to obtain the fused feature map.

9. The rice infection panicle segmentation method based on a progressive fusion network of receptive fields according to claim 1, characterized in that, The segmentation head unit receives the fused multi-level features, gradually restores the spatial resolution through multiple upsampling operations, and performs skip connections and fusion with the features of the corresponding level after each upsampling, finally generating the segmentation mask through a convolutional layer.

Images