Silkworm egg micro-particle virus intelligent detection method

Abstract

The application discloses a kind of silkworm egg microparticle virus intelligent detection methods, comprising S1: collecting the continuous multiple frame microscopic image sequence of silkworm egg sample;S2: generate single-frame clear image after deblurring;S3: output virus candidate frame and its initial confidence;S4: the virus candidate frame and its initial confidence are input to a detection credibility verification module of uncertainty perception and explainability enhancement, and generate gradient weighted class activation mapping heat map, according to the prediction uncertainty variance and image quality dynamic adjustment confidence threshold, output final detection result, this kind of silkworm egg microparticle virus intelligent detection method improves the recovery quality of complex motion blurred image, while, reduce the false detection rate and the missed detection rate of small virus detection, second, provide the uncertainty quantification and explainability output of detection result.

Description

Technical Field

[0001] This invention relates to the field of silkworm seed virus detection, and more particularly to an intelligent detection method for silkworm seed microparticle viruses. Background Technology

[0002] Currently, the main methods for detecting silkworm microparticle viruses include manual microscopic examination, molecular biology methods, and machine vision-based detection methods. Meanwhile, existing technologies use an improved DeblurGANv2 model to deblur motion-blurred images and construct an NN-YOLOv8 detection network based on the YOLOv8 model, which improves the level of automation of detection to a certain extent.

[0003] Existing intelligent detection methods for silkworm microparticle viruses mainly suffer from the following problems: Microparticle viruses are suspended in liquid samples and exhibit irregular, nonlinear, and rapid motion under a microscope, leading to complex spatially variable motion blur in the sampled images. Existing deblurring models (such as DeblurGANv2) are based on convolutional neural networks, which have limited receptive fields and struggle to model global spatiotemporal dependencies. When the blur trajectory is complex, the deblurred image is prone to artifacts, edge ringing effects, and texture distortion, causing viral morphology distortion and directly affecting the accuracy of subsequent detection. Furthermore, microparticle viruses are tiny, occupying only about 15×25 pixels in a microscopic image. Additionally, samples often contain impurities such as silkworm moth fragments, air bubbles, and nematodes, whose morphology and grayscale values ​​are also affected. Because silkworm eggs are highly similar to viruses, existing detection models (such as NN-YOLOv8) use single-scale convolutional kernels and fixed sampling positions, making it difficult to fully extract fine-grained features of tiny targets. Furthermore, they lack a discriminative learning mechanism for distinguishing between viruses and impurities. In complex, densely packed scenarios, the model is prone to misdetecting impurities as viruses or missing detections of adhering or overlapping virus individuals. The detection accuracy cannot meet production-level requirements. Secondly, existing detection methods only output detection boxes and confidence scores, without providing credibility assessments (such as prediction variance and false positive probability) or visual explanations of the decision-making basis. In practical applications of silkworm egg quarantine, if the model outputs incorrect results (false positives leading to the rejection of qualified silkworm eggs, and false negatives leading to the entry of diseased silkworm eggs into production), it will cause significant economic losses. The lack of uncertainty quantification in the "black box" output fails to provide quality inspectors with auxiliary decision-making support, making it difficult to meet the requirements of high credibility and traceability of detection results in the agricultural quarantine field. Summary of the Invention

[0004] The purpose of this invention is to address the shortcomings of existing image deblurring models in adapting to complex motion trajectories, the insufficient ability of target detection models to distinguish between microviruses and impurities, and the lack of uncertainty quantification and interpretability in detection results. This invention provides an intelligent detection method for silkworm microparticle viruses. According to the detection method of this invention, the restoration quality of complex motion-blurred images is improved, while the false detection rate and false negative rate of microvirus detection are reduced. Furthermore, it provides uncertainty quantification and interpretability output for the detection results.

[0005] To achieve the above-mentioned objectives, the technical solution adopted by this invention is as follows: According to one aspect of the present invention, a method for intelligent detection of silkworm seed microparticle viruses is provided, comprising the following steps: S1: A sequence of multiple consecutive microscopic images of silkworm egg samples; S2: Input the continuous multi-frame microscopic image sequence into a generative adversarial deblurring network based on spatiotemporal feature modeling and differentiable kernel estimation, extract the temporal features of the image sequence and estimate the spatially variable motion kernel of each pixel, and generate a deblurred single-frame clear image. S3: Input the deblurred single-frame clear image into a small target detection network based on cross-scale feature collaboration and deformable convolution, extract cross-scale multi-level features and use deformable convolution to adaptively adjust the sampling position, and output the virus candidate box and its initial confidence. S4: Input the virus candidate boxes and their initial confidence scores into a detection confidence verification module that is aware of uncertainty and enhances interpretability. Calculate the prediction uncertainty variance of each detection box using the Monte Carlo Dropout method and generate a gradient-weighted class activation map heatmap. Dynamically adjust the confidence threshold based on the prediction uncertainty variance and image quality, and output the final detection result.

[0006] Preferably, the generative adversarial deblurring network based on spatiotemporal feature modeling and differentiable kernel estimation mentioned in step S2 includes: The temporal feature extraction module uses a 3D convolutional neural network to extract spatiotemporal features from the continuous multi-frame microscopic image sequence; The differentiable motion kernel estimation module regresses from the spatiotemporal features to generate a spatially variable motion kernel for each pixel through a lightweight sub-network; the multi-scale generator takes the spatiotemporal features and the spatially variable motion kernel as conditional inputs to generate a deblurred single-frame clear image. A global-local dual discriminator is used to distinguish between the global realism of the generated image and the local structural integrity of the virus candidate region.

[0007] Preferably, the small target detection network based on cross-scale feature collaboration and deformable convolution in step S3 includes: The cross-scale feature collaboration module adopts a bidirectional cross-scale feature fusion mechanism, which fuses high-resolution spatial features with low-resolution semantic features through learnable weights. The deformable convolutional attention module replaces part of the standard convolution with deformable convolution and combines spatial-channel dual attention mechanism to enhance the feature response of virus regions; The virus-impurity contrastive learning module introduces a contrastive learning loss function during the training phase to distinguish virus features from impurity features in the embedding space.

[0008] Preferably, the fusion process of the cross-scale feature collaboration module is represented as follows: ; in, For the features of layer l, The characteristics after fusion The fusion weights are learnable, with Up and Down representing upsampling and downsampling operations, respectively.

[0009] Preferably, the contrastive loss function of the virus-impurity contrastive learning module is: ; in, Characteristics of viral samples For positive sample features, For impurity negative samples, sim(·) represents the cosine similarity. Let N be the temperature coefficient, and N be the set of negative samples.

[0010] Preferably, the calculation of the prediction uncertainty variance for each detection box using the Monte Carlo Dropout method in step S4 specifically includes: Enable the Dropout layer in the detection head, perform K random forward propagations on the same input image, and record the confidence score for each propagation. Calculate the confidence mean With uncertainty and variance : ; When outputting the detection results, the mean confidence level and the variance of uncertainty are also output simultaneously.

[0011] Preferably, the step S4 of dynamically adjusting the confidence threshold based on the prediction uncertainty variance and image quality includes: Based on the quality evaluation index of the deblurred image and the prediction uncertainty variance, the detection confidence threshold is dynamically adjusted. : ; in, Based on the threshold, Here, Q is the adjustment coefficient, and Q is the image quality evaluation index value. and These are the preset minimum and maximum image quality values.

[0012] Preferably, the generation of the gradient-weighted class activation map heatmap in step S4 specifically includes: For detection box b, its class activation graph ; in For the first Each feature map = For gradient weights, For detection box Category scores.

[0013] Preferably, the method further includes step S5: storing the final detection result in a database, wherein the final detection result includes virus location, virus quantity, confidence mean, uncertainty variance, and the heatmap.

[0014] Preferably, a silkworm seed microparticle virus intelligent detection system includes: The image acquisition module is used to acquire a continuous multi-frame microscopic image sequence of silkworm egg samples; A deblurring module is used to perform the operation described in step S2 of claim 1; The target detection module is used to perform the operation described in step S3 of claim 1; A trusted verification module is used to perform the operation described in step S4 of claim 1; The storage module is used to store the test results.

[0015] In summary, due to the adoption of the above technical solution, the beneficial effects of the present invention are: 1. This invention employs a generative adversarial deblurring network based on spatiotemporal feature modeling and differentiable kernel estimation. It extracts temporal features from multiple consecutive frames of images through a three-dimensional convolutional neural network and estimates the spatially variable motion kernel of each pixel through a differentiable molecular network. This enables accurate modeling of the nonlinear and non-uniform motion trajectory of microparticle viruses. The network structure reduces the number of artifacts and edge ringing effects in the deblurred image and controls the degree of structural distortion of the virus morphology, providing higher quality input images for subsequent detection processes.

[0016] 2. This invention employs a small target detection network based on cross-scale feature collaboration and deformable convolution. It enhances the collaborative expression of multi-level semantic information through a bidirectional cross-scale feature fusion mechanism, and enables the sampling position to be adaptively adjusted according to the irregular shape of the virus through deformable convolution. Furthermore, it explicitly distinguishes between viruses and impurities in the feature embedding space through a contrastive learning module. This network structure enhances the model's ability to identify tiny viruses in scenarios with complex backgrounds and interference from impurities, and reduces the missed detection of adhering and overlapping viruses.

[0017] 3. This invention uses the Monte Carlo Dropout method to calculate the prediction uncertainty variance of each detection box and generates a gradient-weighted class activation map heatmap. At the same time, the confidence threshold is dynamically adjusted according to image quality and uncertainty variance. The output results include the confidence mean, uncertainty variance, and heatmap, enabling quality inspectors to obtain the credibility basis of the model's judgment and a visual interpretation of the decision. This method makes the detection results verifiable and traceable, meeting the requirements of high-reliability detection in the field of agricultural quarantine. Attached Figure Description

[0018] Figure 1 This is a flowchart of the overall method of the present invention; Figure 2 This is a diagram of the deblurring network structure of the present invention; Figure 3 This is a diagram of the small target detection network structure of the present invention; Figure 4 This is a flowchart of the uncertainty quantification and credibility verification process of this invention; Figure 5 This is a structural diagram of the cross-scale feature collaboration module of the present invention; Figure 6 This is a structural diagram of the deformable convolutional attention module of the present invention; Figure 7 This is a schematic diagram of the MC-Dropout uncertainty quantification of the present invention; Figure 8 This is a schematic diagram of the Grad-CAM thermal map of the present invention. Detailed Implementation

[0019] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and preferred embodiments. However, it should be noted that many details listed in the specification are merely to provide the reader with a thorough understanding of one or more aspects of the invention, and these aspects of the invention can be achieved even without these specific details.

[0020] Please see Figures 1 to 8 This invention provides an intelligent detection method for silkworm seed microparticle viruses, the technical solution of which is as follows: This embodiment of the intelligent detection method for silkworm seed microparticle viruses includes the following steps: S1: A sequence of multiple consecutive microscopic images of silkworm egg samples; S2: Input the continuous multi-frame microscopic image sequence into a generative adversarial deblurring network based on spatiotemporal feature modeling and differentiable kernel estimation, extract the temporal features of the image sequence and estimate the spatially variable motion kernel of each pixel, and generate a deblurred single-frame clear image. S3: Input the deblurred single-frame clear image into a small target detection network based on cross-scale feature collaboration and deformable convolution, extract cross-scale multi-level features and use deformable convolution to adaptively adjust the sampling position, and output the virus candidate box and its initial confidence. S4: Input the virus candidate boxes and their initial confidence scores into a detection confidence verification module that is aware of uncertainty and enhances interpretability. Calculate the prediction uncertainty variance of each detection box using the Monte Carlo Dropout method and generate a gradient-weighted class activation map heatmap. Dynamically adjust the confidence threshold based on the prediction uncertainty variance and image quality, and output the final detection result.

[0021] It is worth noting that, specifically, the generative adversarial deblurring network based on spatiotemporal feature modeling and differentiable kernel estimation mentioned in step S2 includes: The temporal feature extraction module uses a 3D convolutional neural network to extract spatiotemporal features from the continuous multi-frame microscopic image sequence; The differentiable motion kernel estimation module regresses from the spatiotemporal features to generate a spatially variable motion kernel for each pixel through a lightweight sub-network; the multi-scale generator takes the spatiotemporal features and the spatially variable motion kernel as conditional inputs to generate a deblurred single-frame clear image. A global-local dual discriminator is used to distinguish between the global realism of the generated image and the local structural integrity of the virus candidate region.

[0022] Next, specifically, the small target detection network based on cross-scale feature collaboration and deformable convolution described in step S3 includes: The cross-scale feature collaboration module adopts a bidirectional cross-scale feature fusion mechanism, which fuses high-resolution spatial features with low-resolution semantic features through learnable weights. The deformable convolutional attention module replaces part of the standard convolution with deformable convolution and combines spatial-channel dual attention mechanism to enhance the feature response of virus regions; The virus-impurity contrastive learning module introduces a contrastive learning loss function during the training phase to distinguish virus features from impurity features in the embedding space.

[0023] Specifically, the fusion process of the cross-scale feature collaboration module is expressed as follows: ; in, For the features of layer l, The characteristics after fusion The fusion weights are learnable, with Up and Down representing upsampling and downsampling operations, respectively.

[0024] Furthermore, specifically, the contrastive loss function of the virus-impurity contrastive learning module is: ; in, Characteristics of viral samples For positive sample features, For impurity negative samples, sim(·) represents the cosine similarity. Let N be the temperature coefficient, and N be the set of negative samples.

[0025] It is worth noting that, specifically, step S4, which involves calculating the prediction uncertainty variance for each detection box using the Monte Carlo Dropout method, specifically includes: Enable the Dropout layer in the detection head, perform K random forward propagations on the same input image, and record the confidence score for each propagation. Calculate the confidence mean With uncertainty and variance : ; When outputting the detection results, the mean confidence level and the variance of uncertainty are also output simultaneously.

[0026] Preferably, specifically, the step S4 of dynamically adjusting the confidence threshold based on the prediction uncertainty variance and image quality includes: Based on the quality evaluation index of the deblurred image and the prediction uncertainty variance, the detection confidence threshold is dynamically adjusted. : ; in, Based on the threshold, Here, Q is the adjustment coefficient, and Q is the image quality evaluation index value. and These are the preset minimum and maximum image quality values.

[0027] Specifically, step S4, generating the gradient-weighted class activation map heatmap, includes: For detection box b, its class activation graph ; in For the first Each feature map = For gradient weights, For detection box Category scores.

[0028] Secondly, specifically, it also includes step S5: storing the final detection results in the database, the final detection results including virus location, virus quantity, confidence mean, uncertainty variance and the heat map.

[0029] Finally, specifically, a silkworm seed microparticle virus intelligent detection system includes: The image acquisition module is used to acquire a continuous multi-frame microscopic image sequence of silkworm egg samples; A deblurring module is used to perform the operation described in step S2 of claim 1; The target detection module is used to perform the operation described in step S3 of claim 1; A trusted verification module is used to perform the operation described in step S4 of claim 1; The storage module is used to store the test results.

[0030] In addition, the circuits, electronic components and modules involved in this invention are all existing technologies, which can be fully implemented by those skilled in the art, and need not be elaborated upon. The content protected by this invention does not involve any improvement to the internal structure and method. Combination Figures 1-8 The specific usage process of the intelligent detection method for silkworm seed microparticle viruses in this embodiment is as follows: Example 1 A smart detection method for silkworm seed microparticle viruses, the specific steps of which are as follows: Step 1: Acquire image sequence. Place the female moth sample to be tested on a glass slide, add deionized water, cover with a coverslip, and place on the electron microscope stage. Adjust the microscope to the preset magnification and continuously acquire multiple frames of microscopic images at the set sampling rate. The image format is bitmap.

[0031] Step two, deblurring. The multi-frame image sequence acquired in Step one is input into a pre-trained generative adversarial deblurring network based on spatiotemporal feature modeling and differentiable kernel estimation. This network first extracts the temporal features of the image sequence through a 3D convolutional neural network, with the convolutional kernel size and stride set according to preset parameters. Subsequently, a differentiable motion kernel estimation sub-network regresses from the temporal feature map to generate a spatially variable motion kernel for each pixel. A multi-scale generator receives the original blurred image, temporal features, and motion kernel, and generates a single-frame clear image after deblurring through an encoder-decoder structure. The global discriminator and local discriminator are adversarially trained to assess the realism of the generated image.

[0032] Step 3, Small Object Detection. The cleared image generated in Step 2 is input into a pre-trained small object detection network based on cross-scale feature collaboration and deformable convolution. This network uses the YOLOv8 series models as a baseline framework, adding a cross-scale feature collaboration module at the end of the backbone network, replacing some standard convolutions with deformable convolutions in the neck network, and introducing virus-impurity contrastive learning loss during the training phase. The network outputs virus candidate boxes and their initial confidence scores.

[0033] Step 4: Uncertainty Quantification and Confidence Verification. Each detection box and its confidence score from Step 3 are input into the uncertainty perception module. A Dropout layer is enabled in the detection head, performing multiple random forward propagations on the same image, recording the confidence score for each output, and calculating the mean and variance of the confidence scores. Simultaneously, a gradient-weighted activation mapping method is used to generate a heatmap for each detection box, displaying the image region the model focuses on during decision-making. Based on the image quality evaluation index and uncertainty variance of the deblurred image, the confidence threshold is dynamically adjusted: when the uncertainty variance exceeds a preset upper limit, the confidence threshold is increased; when the image quality falls below a preset lower limit, the confidence threshold is further increased. The final output is the detection result after dynamic threshold filtering, including virus location coordinates, virus count, mean confidence score, uncertainty variance, and a heatmap.

[0034] Step 5: Store the detection results. Store all the detection information output in Step 4 into the database. Each record includes the detection time, image file name, number of viruses, location coordinates of each virus, mean and variance of confidence score, and path to the heatmap file. Example 2 This embodiment is basically the same as Embodiment 1, except for the following parameter settings: In step one, the number of frames continuously acquired is set to a first preset value.

[0035] In step two, the kernel size of the 3D convolutional neural network is set to be larger than that in Example 1, the number of output channels is set to be larger than that in Example 1, and the motion kernel size is set to be larger than that in Example 1. The learning rates of the generator and discriminator are set to values ​​different from those in Example 1, and the number of training epochs is set to values ​​different from those in Example 1.

[0036] In step three, the input image size is set to be larger than that in Example 1, the initial detection confidence threshold is set to a value lower than that in Example 1, and the cross-union ratio (CUI) threshold for non-maximum suppression is set to a value lower than that in Example 1. The temperature coefficient in the contrastive learning loss function is set to a preset value, and the size of the negative sample set is set to a preset value.

[0037] In step four, the Dropout probability is set to a value higher than that in Example 1, and the number of Monte Carlo forward propagations is set to a value greater than that in Example 1. The first and second adjustment coefficients in the dynamic threshold adjustment formula are set to preset values. The image quality evaluation index uses the structural similarity index; when this index is lower than a preset threshold, the confidence threshold is increased by a preset percentage.

[0038] This embodiment is applicable to scenarios where the virus density in the sample is high and the interference from impurities is severe. By increasing the input resolution and adjusting the detection threshold, the false negative rate can be further reduced. Example 3 This embodiment provides a lightweight implementation method suitable for real-time detection of embedded devices or mobile terminals.

[0039] In step one, the number of consecutively acquired frames is set to be less than the value in Example 1, the sampling rate is set to be less than the value in Example 1, and the image resolution is set to be less than the value in Example 1.

[0040] In step two, the three-dimensional convolutional neural network in the generative adversarial deblurring network based on spatiotemporal feature modeling and differentiable kernel estimation is replaced with a lightweight convolutional long short-term memory network module to reduce the number of parameters. The motion kernel size is set to be smaller than the value in Example 1. The number of channels in the convolutional layer in the generator network is halved, and the global discriminator is replaced with a single-scale discriminator.

[0041] In step three, the small object detection network is replaced with a lightweight version instead of the standard-sized model. The number of fusion layers in the cross-scale feature collaboration module is set to less than the value in Example 1, and deformable convolutions are only retained in the last feature map. The contrastive learning module is frozen during embedded deployment, using only pre-trained weights for inference.

[0042] In step four, the number of Monte Carlo forward propagations is set to be less than that in Example 1, and the dynamic threshold adjustment is based solely on the uncertainty variance, without relying on image quality evaluation. Heatmap generation is set as an optional function, and users can choose whether to enable it through the interactive interface.

[0043] This embodiment is applicable to handheld testing equipment used in silkworm seed production sites, which can achieve real-time testing while ensuring basic testing accuracy.

[0044] The above description is only a preferred embodiment of the present invention. It should be noted that for those skilled in the art, several improvements and modifications can be made without departing from the principle of the present invention, and these improvements and modifications should also be considered within the scope of protection of the present invention.

Claims

1. A method for intelligent detection of silkworm seed microparticle viruses, characterized in that, Includes the following steps: S1: A sequence of multiple consecutive microscopic images of silkworm egg samples; S2: Input the continuous multi-frame microscopic image sequence into a generative adversarial deblurring network based on spatiotemporal feature modeling and differentiable kernel estimation, extract the temporal features of the image sequence and estimate the spatially variable motion kernel of each pixel, and generate a deblurred single-frame clear image. S3: Input the deblurred single-frame clear image into a small target detection network based on cross-scale feature collaboration and deformable convolution, extract cross-scale multi-level features and use deformable convolution to adaptively adjust the sampling position, and output the virus candidate box and its initial confidence. S4: Input the virus candidate boxes and their initial confidence scores into a detection confidence verification module that is aware of uncertainty and enhances interpretability. Calculate the prediction uncertainty variance of each detection box using the Monte Carlo Dropout method and generate a gradient-weighted class activation map heatmap. Dynamically adjust the confidence threshold based on the prediction uncertainty variance and image quality, and output the final detection result.

2. The intelligent detection method for silkworm microparticle viruses according to claim 1, characterized in that: The generative adversarial deblurring network based on spatiotemporal feature modeling and differentiable kernel estimation mentioned in step S2 includes: The temporal feature extraction module uses a 3D convolutional neural network to extract spatiotemporal features from the continuous multi-frame microscopic image sequence; The differentiable motion kernel estimation module regresses from the spatiotemporal features to generate a spatially variable motion kernel for each pixel through a lightweight sub-network; the multi-scale generator takes the spatiotemporal features and the spatially variable motion kernel as conditional inputs to generate a deblurred single-frame clear image. A global-local dual discriminator is used to distinguish between the global realism of the generated image and the local structural integrity of the virus candidate region.

3. The intelligent detection method for silkworm microparticle viruses according to claim 1, characterized in that: The small target detection network based on cross-scale feature collaboration and deformable convolution mentioned in step S3 includes: The cross-scale feature collaboration module adopts a bidirectional cross-scale feature fusion mechanism, which fuses high-resolution spatial features with low-resolution semantic features through learnable weights. The deformable convolutional attention module replaces part of the standard convolution with deformable convolution and combines spatial-channel dual attention mechanism to enhance the feature response of virus regions; The virus-impurity contrastive learning module introduces a contrastive learning loss function during the training phase to distinguish virus features from impurity features in the embedding space.

4. The intelligent detection method for silkworm microparticle viruses according to claim 3, characterized in that, The fusion process of the cross-scale feature collaboration module is represented as follows: ; in, For the features of layer l, The characteristics after fusion The fusion weights are learnable, with Up and Down representing upsampling and downsampling operations, respectively.

5. The intelligent detection method for silkworm microparticle viruses according to claim 3, characterized in that: The contrastive loss function of the virus-impurity contrastive learning module is: ; in, Characteristics of viral samples For positive sample features, For impurity negative samples, sim(·) represents the cosine similarity. Let N be the temperature coefficient, and N be the set of negative samples.

6. The intelligent detection method for silkworm microparticle viruses according to claim 1, characterized in that, Step S4, which involves calculating the prediction uncertainty variance for each detection box using the Monte Carlo Dropout method, specifically includes: Enable the Dropout layer in the detection head, perform K random forward propagations on the same input image, and record the confidence score for each propagation. Calculate the confidence mean With uncertainty and variance : ； When outputting the detection results, the mean confidence level and the variance of uncertainty are also output simultaneously.

7. The intelligent detection method for silkworm seed microparticle viruses according to claim 1, characterized in that, The step S4, which involves dynamically adjusting the confidence threshold based on the prediction uncertainty variance and image quality, includes: Based on the quality evaluation index of the deblurred image and the prediction uncertainty variance, the detection confidence threshold is dynamically adjusted. : ; in, Based on the threshold, Here, Q is the adjustment coefficient, and Q is the image quality evaluation index value. and These are the preset minimum and maximum image quality values.

8. The intelligent detection method for silkworm seed microparticle viruses according to claim 1, characterized in that, The generation of gradient-weighted class activation map heatmap in step S4 specifically includes: For detection box b, its class activation graph ; in For the first Each feature map = For gradient weights, For detection box Category score.

9. The intelligent detection method for silkworm seed microparticle viruses according to claim 1, characterized in that, It also includes step S5: storing the final detection result in the database, the final detection result including virus location, virus quantity, confidence mean, uncertainty variance and the heat map.

10. The intelligent detection method for silkworm seed microparticle viruses according to claim 1, characterized in that, A smart detection system for silkworm microparticle viruses includes: The image acquisition module is used to acquire a continuous multi-frame microscopic image sequence of silkworm egg samples; A deblurring module is used to perform the operation described in step S2 of claim 1; The target detection module is used to perform the operation described in step S3 of claim 1; A trusted verification module is used to perform the operation described in step S4 of claim 1; The storage module is used to store the test results.

Images