AI recognition-based silkworm microspore spore rapid detection method and system

Abstract

The application discloses a kind of based on AI identification's silkworm microspore spore rapid detection method and system, RCL-YOLO realizes the effective promotion of detection precision by three collaborative improvements;RFCBAMConv_C3k2 module is introduced in backbone network, and the capture ability to weak feature of spore is enhanced;CAFMFusion module is introduced in neck network, and the fine fusion of shallow details and deep semantics is realized;LQEHead module is introduced in detection head, and low-quality prediction frame is inhibited by positioning quality calibration;On self-built image dataset, the mAP50 and mAP50:95 of RCL-YOLO reach 85.38% and 32.96% respectively, compared with original YOLOv11n, increase by 2.94 and 0.98 percentage points respectively, the precision and recall rate increase by 3.79 and 2.51 percentage points respectively, and model memory occupancy is only 7.5MB.

Description

Technical Field

[0001] This invention relates to the field of detection model technology, specifically to a rapid detection method and system for silkworm microspores based on AI recognition. Background Technology

[0002] Silkworm microsporidiasis is a devastating infectious disease caused by Nosema bombycis, and is listed as the only legally required quarantine object in my country's sericulture industry. Microscopic examination, as the main detection method of grassroots quality inspection agencies, has shortcomings such as limited detection sensitivity and reliance on the experience of inspectors, making it difficult to meet the needs of rapid diagnosis of early low-concentration infections.

[0003] Silkworm microsporidiasis is one of the most serious pathogens threatening sericulture. Caused by the silkworm microsporidia (Nosema bombycis), it is a devastating infectious disease and is listed as the only legally mandated quarantine target for sericulture in my country. This pathogen can be transmitted through ingestion and vertical transmission within the egg, exhibiting genetic inheritance. During the incubation period, silkworms show no obvious external symptoms. Once an outbreak occurs, it can lead to widespread disease in silkworm populations, a sharp drop in cocoon production of over 30%, or even total crop failure, and in severe cases, can destroy the entire silkworm rearing system. Therefore, developing highly sensitive, highly specific, and readily applicable early rapid diagnostic technologies has become an urgent engineering and technical requirement for ensuring the sustainable development of the sericulture industry.

[0004] Microsporidia spores are only 2-4 μm in diameter, making them extremely small and difficult to detect through routine microscopic examination in the early stages of low-concentration infection (≤10³ spores / mL). While molecular detection techniques have advantages in sensitivity and early diagnosis, microscopic examination remains the most economical method due to limitations in equipment and operating costs at grassroots silkworm breeding farms. Current research summarizes the principles, advantages, and disadvantages of molecular detection techniques such as PCR, qPCR, and LAMP for silkworm microsporidia, pointing out that although molecular methods have significant advantages in early diagnosis, their widespread application still faces practical obstacles such as high equipment costs and complex operations. Microscopic examination, on the other hand, uses an optical microscope to directly observe microsporidia in samples, requiring no expensive instruments or complex reagents. It can be routinely used by grassroots silkworm breeding farms and quality inspection institutions, offering low detection costs and easy-to-master operation techniques. According to the "Silkworm Seed Management Regulations" and the "Silkworm Original Seed Inspection Procedures" (NY / T...), this method is suitable for silkworm breeding farms and quality inspection institutions. Standards such as 327-1997 stipulate that microscopic examination of female moths remains the legally mandated method for silkworm seed quality inspection in my country. Each batch of silkworm seeds must be sampled for microscopic examination of female moths to ensure that the pass rate for silkworm seed microsporidiosis meets the standards. Although this method has shortcomings such as limited detection sensitivity and reliance on the experience of inspection personnel, it remains the preferred method for silkworm seed quality inspection in my country and even developing countries due to its advantages such as low cost, high accessibility, and intuitive and reliable results. It plays an irreplaceable and fundamental role in ensuring the safety of silkworm seeds and the healthy development of the industry.

[0005] In recent years, automated detection systems based on machine vision have shown great potential in agricultural pest and disease monitoring. Some scholars have applied the YOLO algorithm to plant and animal phenotypic research; others have constructed multi-scale residual networks fused with SENet, achieving an accuracy of 98.72%; still others have proposed RGB color component feature extraction methods based on data density, providing an effective means for silkworm image segmentation; lightweight dense multi-scale attention networks have been used for wheat leaf rust identification; and for spore detection in microscopic images, a YOLO-ASF-MobileViT model based on YOLOv5 has been proposed, effectively improving the detection accuracy of microspores under densely adhered conditions; a GCS-YOLOv8 network has been proposed, achieving an accuracy of 92.6% in cucumber pathogen spore detection; and improvements have been made to YOLO... A lightweight method for detecting rice blast spores using v8 was proposed; in addition, a lightweight silkworm identification method based on multi-scale feature fusion was proposed, achieving high-precision real-time detection in dense silkworm scenes; a YOLO model based on an improved attention mechanism was used to achieve rapid identification of greenhouse tomatoes; the ability to detect targets in complex environments was improved through deep feature fusion of Transformer-CNN; and a YOLOv10s-MAD model based on Manhattan distance attention and gating dual branches was designed, achieving a mAP50 of 92.2% in the complex identification of multiple diseases on mulberry leaves. The above studies have verified the effectiveness of attention mechanisms, multi-scale fusion and other strategies in small target detection, providing direct reference for subsequent proposed solutions.

[0006] Although the above studies have promoted the development of intelligent diagnosis of silkworm diseases and small target detection technology, the "weak features" of silkworm microspores have not yet been studied. The so-called weak features refer to the fact that the spores are extremely small in size (only 2~4μm), the gray transition between the edge and the background is smooth, the contrast is low, and they are often highly similar in texture to translucent interference objects such as impurities and bubbles, which makes them appear as "weak signal and strong noise" in microscopic images.

[0007] The existing model has the following three shortcomings: First, it lacks a signal-to-noise ratio enhancement mechanism for weak features, making it difficult to effectively separate spores from impurities in complex backgrounds; second, it lacks modeling for the positioning uncertainty of regions with blurred boundaries, leading to prediction box offset; third, it lacks a refined fusion strategy for shallow details and deep semantics, making it easy for weak features to be lost in cross-layer transmission. Therefore, a dedicated detection method for weak features of silkworm microspores is urgently needed. Summary of the Invention

[0008] To achieve the above objectives, the present invention provides the following technical solution:

[0009] A rapid detection method for silkworm microspores based on AI recognition, the method comprising:

[0010] Obtain the image dataset of the target spores and perform data preprocessing;

[0011] Based on the existing YOLOv11n framework, an RCL-YOLO model suitable for detecting weak spore features during microscopic examination is constructed:

[0012] In the backbone network, the RFCBAMConv_C3k2 module, which integrates receptive field adaptation and channel space dual attention, is introduced to replace the original C3k2 structure. In the neck feature fusion stage, the context-aware dual-branch attention fusion module CAFMFusion is introduced to replace the original operation. Through the collaboration of local convolution and global self-attention, as well as pixel-level weighting, it achieves the fusion of shallow spatial details and deep semantic information. In the detection head, the localization quality estimation head LQEHead is introduced to explicitly evaluate the localization accuracy of the bounding box and calibrate the classification confidence.

[0013] Furthermore, the image dataset was acquired by taking pictures with an optical microscope, using the microscope and its accompanying camera, and the image dataset was taken of the target spores under different lighting and density conditions.

[0014] Furthermore, when preprocessing the image dataset, the enhancement strategies employed include at least: horizontal and vertical flipping, small-angle random rotation, scaling, and random brightness; where the small angle is a preset angle that can be adjusted according to the actual situation or pre-set conditions.

[0015] Furthermore, the RFCBAMConv_C3k2 module replaces the standard convolution within the Bottleneck in the original C3k2 with RFCBAMConv; where RFCBAMConv is derived by fusing CBAM improvements on top of RFAConv; where CBAM is a concatenated mechanism of channel attention and spatial attention.

[0016] Furthermore, in CBAM, channel attention (CA) is used to improve the signal-to-noise ratio of weak features. It extracts multi-receptor field features in parallel through grouped convolutions of different scales, and generates channel weights through spatial interaction matrix and Softmax normalization to highlight the weak spore signals in the channel dimension. Spatial attention (SA) is used to locate weak feature regions. The feature map after channel attention weighting is average pooled and max pooled along the channel dimension, and after concatenation, a spatial weight map is generated through 1×1 convolution and Sigmoid activation function to mark the high-response region where the spore is located.

[0017] Furthermore, RFCBAMConv simulates the receptive field mechanism of the human visual system, introducing the ability to adaptively adjust the receptive field in spatial attention through multi-scale grouped convolution. RFCBAMConv dynamically adjusts the feature response intensity at different spatial locations through grouped convolution, which is equivalent to applying a non-uniform receptive field weight distribution in different regions, enabling the network to flexibly focus according to the spore scale and morphology, and adapt to the weak feature morphological changes of spores of different sizes.

[0018] The improved RFCBAMConv_C3k2 module retains residual connections through a dual-path design;

[0019] In the backbone and neck networks, the original C3k2 modules in layers 2, 4, 6, 8, 13, 16, 19, and 22 are replaced with RFCBAMConv_C3k2 to form a hierarchical weak feature enhancement mechanism. The shallow modules P2 and P3 enhance the spore edge contour through spatial attention, initially locking candidate regions in complex backgrounds and preserving the spatial structure of weak features. The deep modules P4 and P5 use channel attention and receptive field adaptation to select high-confidence features at the semantic level to enhance the semantic discriminative power of weak features.

[0020] Furthermore, the design of the dual-branch attention fusion module CAFMFusion consists of two steps:

[0021] The first step is the convolution and attention fusion module CAFM; firstly, shallow features and deep features are added element-wise to obtain initial fused features, which are then fed into two branches in parallel:

[0022] In the local branch, 1×1 convolution and Channel Shuffle operation are applied sequentially to enhance channel interaction, and then 3×3 convolution is used to capture local spatial correlation.

[0023] In the global branch, after dimensionality reduction using 1×1 convolution, a multi-head self-attention mechanism is used to generate queries, keys, and values. The values ​​are weighted and aggregated through an attention map to capture long-range dependencies across spatial locations. The outputs of the two branches are connected and integrated through convolution to achieve initial multi-scale, cross-level feature fusion.

[0024] The second step is pixel attention weighting; the enhanced features from the CAFM module are connected along the channel dimension to the initial fused features, and then input into the grouped convolution, i.e., 7×7 convolution kernel, with the number of groups equal to the number of channels. A pixel-level weight map is generated through the Sigmoid activation function. This weight map is multiplied by the input shallow features and deep features respectively, and spatial importance coefficients are adaptively assigned to both to complete fine-grained spatial information fusion.

[0025] Finally, a 1×1 convolution is performed to output multi-scale fused features;

[0026] In the RCL-YOLO neck structure, the dual-branch attention fusion module CAFMFusion is deployed at the key cross-layer connection between the backbone and the head.

[0027] Furthermore, the localization quality estimation head LQEHead predicts localization quality scores by utilizing bounding box distribution statistics and performs additive calibration on classification scores, employing the core logic of the distribution-guided quality predictor in GFLV2.

[0028] The complete structure of LQEHead includes: Detect_LQE-Branch, which is an improved detection head structure. Based on the original structure, it fuses the output of the classification branch with the output of a newly added LQE module. The input of the LQE module comes from the features of the bounding box regression branch. LQE is the core module of LQEHead, showing its internal statistical feature extraction and quality estimation MLP steps. It receives the predicted corner distribution vector from the bounding box regression branch in Detect_LQE-Branch, extracts features reflecting localization uncertainty through TopK statistics and Mean pooling, and maps them to the final localization quality adjustment value through the MLP network.

[0029] Furthermore, the LQEHead calculation process consists of three steps:

[0030] The first step is TopK statistical feature extraction. For the four-sided distribution vector output by the bounding box regression branch, the first four maximum values ​​and their mean are extracted for each side, resulting in 5-dimensional statistical features for each side, for a total of 20 dimensions for the four sides.

[0031] The second step is Softmax normalization, which performs Softmax normalization on the extracted statistical features along the channel dimension, converting the feature values ​​into a probability distribution form.

[0032] The third step is MLP mapping, which feeds the normalized 20-dimensional statistical features into a lightweight MLP sub-network and performs feature transformation through two pointwise convolutional layers. The first convolutional layer maps the number of channels from 20 dimensions to the dimensions of the intermediate hidden layer, and performs a non-linear transformation using the ReLU activation function. The second convolutional layer maps the output of the hidden layer to a scalar to generate a localization quality adjustment value. Finally, LQEHead uses an additive calibration mechanism to add the localization quality adjustment value to the initial classification score output by the classification branch to obtain the calibrated final classification score.

[0033] A rapid detection system for silkworm microspores based on AI recognition, the system comprising:

[0034] Data acquisition module: Acquires image datasets of target spores and performs data preprocessing;

[0035] Model building module: Based on the existing YOLOv11n framework, an RCL-YOLO model suitable for detecting weak spore features during microscopic examination is constructed.

[0036] In the backbone network, the RFCBAMConv_C3k2 module, which integrates receptive field adaptation and channel space dual attention, is introduced to replace the original C3k2 structure. In the neck feature fusion stage, the context-aware dual-branch attention fusion module CAFMFusion is introduced to replace the original operation. Through the collaboration of local convolution and global self-attention, as well as pixel-level weighting, it achieves the fusion of shallow spatial details and deep semantic information. In the detection head, the localization quality estimation head LQEHead is introduced to explicitly evaluate the localization accuracy of the bounding box and calibrate the classification confidence.

[0037] This invention provides a rapid detection method and system for silkworm microspores based on AI recognition, which has the following beneficial effects:

[0038] To address the challenge of detecting silkworm microspores in microscopic images due to their extremely small size (2-4 μm), blurred edges, and similar grayscale to background impurities, this application proposes an RCL-YOLO spore detection model based on an improved YOLOv11n. Compared to traditional microscopic methods and existing deep learning models, this model focuses on three core challenges of weak features—insufficient weak feature expression, low efficiency of multi-scale fusion, and uncertain bounding box localization—and implements three improvement measures: an adaptive dual attention module for the backbone receptive field, a neck context-aware feature fusion module, and a detection head localization quality estimation module. This forms a complete weak feature detection scheme from feature extraction and cross-layer fusion to prediction box calibration.

[0039] Specifically, the following is explained: First, the RFCBAMConv_C3k2 module is designed in the backbone network. Through the receptive field adaptive mechanism and channel space dual attention, it is specifically enhanced to capture weak features of spores (edge ​​details, low contrast areas) and improve noise resistance in complex backgrounds.

[0040] Secondly, in the neck feature fusion stage, a context-aware dual-branch attention fusion module CAFMFusion is introduced. Through the collaboration of local convolution and global self-attention, as well as pixel-level weighting, it achieves a refined fusion of shallow weak feature details and deep semantic information, avoiding the loss of weak features in cross-layer transmission.

[0041] Finally, a localization quality estimation head (LQEHead) is introduced into the detection head section. Addressing the localization uncertainty caused by spore boundary ambiguity, it explicitly evaluates the localization accuracy of the bounding boxes and calibrates the classification confidence, effectively suppressing low-quality predicted boxes.

[0042] Experimental validation was conducted on a self-built dataset of microscopic images of silkworm microspores. The results showed that the RCL-YOLO model proposed in this application achieved a detection accuracy of 85.38% mAP50 and 32.96% mAP50:95, which are 2.94 and 0.98 percentage points higher than the original YOLOv11n, respectively. The precision and recall rates were improved by 3.79 and 2.51 percentage points, respectively. Meanwhile, the model's memory footprint is only 7.5 MB. While maintaining a lightweight design, it improves the detection performance of weak features of microspores and can provide an effective technical reference for the intelligent inspection of silkworm seed quality. Attached Figure Description

[0043] Figure 1 This is a schematic diagram of microscopic examination of microspores in this invention;

[0044] Figure 2 This is a data-enhanced image of microspores in this invention;

[0045] Figure 3 This is a diagram of the existing, unmodified YOLO11 structure.

[0046] Figure 4 This is a structural diagram of the improved RCL-YOLO in this invention;

[0047] Figure 5 This is a structural diagram of the RFCBAMConv_C3k2 module in this invention;

[0048] Figure 6 This is a structural diagram of CAFMFusion in this invention;

[0049] Figure 7 This is a structural diagram of the LQEHead in this invention;

[0050] Figure 8 This is a visualization of the ablation experiment results in this invention;

[0051] Figure 9 This is a schematic diagram of the P, R, and map50 curves for different detection models in this invention.

[0052] Figure 10 This is a visualization of the detection results of multiple datasets at different resolutions in this invention;

[0053] Figure 11 is a schematic diagram of the image detection, video detection, and real-time monitoring modes in this invention. Detailed Implementation

[0054] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of the present invention.

[0055] This embodiment provides a rapid detection method for silkworm microspores based on AI recognition. The specific process of this method is described below:

[0056] 1. Image acquisition and data preprocessing:

[0057] 1.1 Data Collection:

[0058] The identification of microsporids mainly relies on microscopic examination techniques. The basis for this method is the unique morphological characteristics of the pathogen under an optical microscope (usually 400-600 times). Microsporids are about 3-5 micrometers in size, oval or elliptical in shape, and have strong refractive properties and a bright appearance.

[0059] The dataset used in this application was photographed in April 2026 under an optical microscope (Nikon) at the Sericulture Technology Extension Station, No. 10 Xiajun Road, Xixiangtang District, Nanning City, Guangxi Zhuang Autonomous Region. Imaging was achieved using the Nikon microscope and its accompanying DigitalSight camera. To obtain diverse data, spores were photographed under different lighting and density conditions to ensure data reliability. Figure 1 The image shows the microscopic examination results of microspores at different densities. Figure 1 The right side of Figures (a) and (b) corresponds to magnified versions of the spores; among them, Figure 1 (a) shows a sparse distribution. The only spore in the image is marked with an arrow. For closer observation, a magnified version is placed... Figure 1 (a) Right side; Figure 1 (b) shows a dense distribution with approximately 80 spores in the image. This data image has been authorized for use by the promotion station and has undergone image screening. After removing samples with irregular spore shapes, a total of 2,500 data images were retained.

[0060] 1.2 Data Preprocessing:

[0061] To enhance the diversity of the dataset and improve the robustness and generalization ability of the research model, this application employs various data augmentation methods. Considering that microspores are small targets, the augmentation strategies, without destroying spore characteristics, include horizontal and vertical flipping, small-angle random rotation, scaling, and random brightness adjustments. Figure 2The images shown are different effects of data enhancement on microspore images containing insufficiently ground impurities obtained during normal microscopic examination at the Sericulture Institute. After data enhancement, the total number of images was expanded to 3500.

[0062] Based on this, invalid samples with excessive blur and those that did not contain the target individuals were further removed, ultimately retaining a total of 3200 valid images. Subsequently, the dataset was randomly divided into stratified subsets in a 7:2:1 ratio. This partitioning method effectively ensured that samples of different densities and shooting distances were evenly distributed among the subsets, providing a reliable data foundation for model training and evaluation. It also helped to alleviate overfitting and improve the model's adaptability in real-world scenarios. The microsporidiosis spores were labeled with vector quadrilaterals using Labelimg and saved locally as TXT files, containing information on the category and disease location.

[0063] 2. Model Basics and Improvements:

[0064] YOLO11 is a next-generation object detection model released by Ultralytics in September 2024. Its network structure is as follows: Figure 3 As shown, it consists of three parts: the backbone network, the neck network, and the detection head;

[0065] The backbone network, based on the CSPDarknet53 architecture, replaces the original C2f module with a C3k2 module. It enhances multi-scale feature extraction capabilities by dynamically adjusting the number of Bottlenecks and convolutional kernel parameters, and introduces a C2PSA spatial awareness module at the end to fuse channel and spatial dimension feature information through a pyramid slicing attention mechanism. The neck network adopts an improved PAN-FPN structure, which achieves cross-layer fusion of shallow details and deep semantics through bidirectional paths from top to bottom and bottom to top. The detection head adopts a decoupled design, with classification and localization branches independent of each other, outputting detection results at three scales: P3, P4, and P5, respectively. Although YOLO11 performs well in general object detection tasks, there is still room for optimization in feature extraction and detection accuracy for microscopic spore targets with extremely small size and blurred boundaries.

[0066] 2.1 Construction of an RCL-YOLO model suitable for detecting weak spore features during microscopic examination:

[0067] To address the challenge of weak feature detection in microscopic images of silkworm microspores, which are characterized by extremely small target size, blurred boundaries, and severe background interference, this application proposes an RCL-YOLO model based on the YOLOv11n framework, with improvements made to the backbone network, neck feature fusion, and detection head. First, the RFCBAMConv_C3k2 module, which integrates adaptive receptive field and dual channel space attention, is introduced into the backbone network to replace the original C3k2 structure, thereby enhancing the network's ability to capture details at the edges of microspores and its noise resistance in complex backgrounds.

[0068] Secondly, in the neck feature fusion stage, a context-aware dual-branch attention fusion module CAFMFusion is introduced to replace the original simple splicing operation. Through the collaboration of local convolution and global self-attention as well as pixel-level weighting, it achieves a refined fusion of shallow spatial details and deep semantic information.

[0069] Finally, a localization quality estimation head (LQEHead) is introduced into the detection head section. By explicitly evaluating the localization accuracy of the bounding box and calibrating the classification confidence, the low-quality predicted boxes caused by spore boundary blurring are effectively suppressed.

[0070] The aforementioned improvements work synergistically to enhance RCL-YOLO's ability to detect weak features of silkworm microspores while maintaining the model's lightweight design. The overall structure of the improved RCL-YOLO is as follows: Figure 4 As shown in the figure, the specific locations of each improved module are marked with red boxes to clearly understand the improvements made to the model.

[0071] 2.2, Adaptive Dual Attention Module for Enhanced Weak Feature Extraction:

[0072] In the microscopic images of silkworm microspores, the target size is extremely small (only 2~4μm in diameter) and is often transparent or semi-transparent. It is highly similar to background impurities, bubbles and other interfering objects in terms of grayscale and texture. During the microscopic imaging process, there are problems such as uneven illumination and focal length deviation, which further exacerbates the difficulty of distinguishing the foreground from the background.

[0073] The above factors cause spores to exhibit typical "weak feature" characteristics, including extremely low pixel ratio, low edge contrast, weak signal strength, and easy to be submerged by background noise. The C3k2 module in the YOLOv11n backbone network uses standard convolution and fixed channel configuration to construct residual blocks, and its feature extraction process has three shortcomings.

[0074] First, the receptive field shape and scale of standard convolution are fixed, making it difficult to adapt to the varied morphology and scale of spores, especially in the early stages of low-concentration infection, where weak foreground signals are easily ignored. Second, traditional convolution lacks the ability to focus on key spatial locations, and spore edge features are easily submerged by background noise in deeper layers. Third, although residual connections alleviate gradient vanishing, they do not distinguish the importance of features in channels and spatial dimensions, resulting in feature redundancy.

[0075] The aforementioned issues result in insufficient discriminative power of the spore features output by the backbone network, leading to even weaker responses to targets with weak features, which directly affects the accuracy of subsequent detection.

[0076] To address the issues of insufficient expression of weak spore features and low signal-to-noise ratio, the RCL-YOLO model proposed in this application introduces an improved module that integrates residual features and receptive field space dual attention into the backbone network, namely RFCBAMConv_C3k2 (C3k2 with Residual Feature and Channel Block Attention ModuleConvolution).

[0077] This module primarily replaces the standard convolution within Bottleneck in the original C3k2 with RFCBAMConv;

[0078] Among them, RFCBAMConv is an improvement on RFAConv by integrating CBAM (channel attention + spatial attention). The original RFAConv effectively solved the problem of insufficient modeling of spatial feature differences within the receptive field of the convolution kernel by introducing a receptive field attention mechanism, and improved network performance with almost no increase in computational overhead. However, the core contribution of RFAConv is concentrated on the dynamic allocation of feature weights in the spatial dimension of the receptive field, and its ability to model attention in the channel dimension is relatively weak, making it difficult to fully focus on the key difference information of weak feature targets in different channels.

[0079] To address the aforementioned shortcomings, this application further introduces a concatenated mechanism of channel attention and spatial attention, namely the RFCBAMConv_C3k2 module, based on RFAConv. This module enhances the feature selection capability of the channel dimension while retaining the original spatial feature enhancement capability of the receptive field through adaptive adjustment of the receptive field and joint modeling of dual attention, thereby improving the network's ability to extract discriminative features of small targets.

[0080] The structure of the RFCBAMConv_C3k2 module is as follows: Figure 5As shown, the part marked in red is the replacement part and the specific structure of the replaced DysnakeConv_C3k2; its core improvement is reflected in two aspects;

[0081] First, a mechanism is introduced that connects channel attention (CA) and spatial attention (SA);

[0082] Channel attention (CA) aims to improve the signal-to-noise ratio of weak features. It extracts multi-receptor field features in parallel through grouped convolutions at different scales, and generates channel weights through spatial interaction matrix and Softmax normalization. This enhances channels sensitive to key features such as spore edges and textures, while suppressing redundant responses caused by uneven illumination and impurities, allowing weak spore signals to stand out along the channel dimension. Spatial attention (SA) aims to accurately locate weak feature regions. It performs average pooling and max pooling along the channel dimension on the feature map weighted by channel attention, concatenates the two maps, and generates a spatial weight map through 1×1 convolution and a sigmoid activation function. This accurately identifies the high-response regions where spores are located, effectively suppressing interference from background regions on weak features. The calculation processes for both are as follows:

[0083] The calculation basis for channel attention CA is as follows:

[0084] (1)

[0085] (2)

[0086] (3)

[0087] (4)

[0088] The calculation basis for spatial attention (SA) is as follows:

[0089] (5)

[0090] (6)

[0091] (7)

[0092] In the Channel Attention (CA) stage, the input feature map is first subjected to parallel convolution operations with kernel sizes of 3 and 5 to extract multi-scale receptive field features, where F represents the original input feature map with dimensions C×H×W (representing the number of channels, height, and width, respectively), Conv3 and Conv5 represent convolution operations with kernel sizes of 3 and 5, respectively, and ϕ represents the fused feature after element-wise addition of the two receptive field features. Subsequently, the fused feature map is compressed into channel descriptors g using the global average pooling operation GAP(ϕ). Next, by constructing a spatial interaction matrix, the row vectors g of the descriptors are used...row and column vector g col Matrix multiplication and element-wise addition are performed, and then normalized using the Softmax function to obtain the channel attention weight vector δ. Finally, the channel attention weights δ are multiplied element-wise along the channel dimension by the original fused feature ϕ to obtain the feature map Δ enhanced by channel attention. CA .

[0093] In the spatial attention (SA) stage, the feature map Δ output from the previous stage... CA As input, average pooling (AvgPool) and max pooling (MaxPool) are performed along the channel dimension to obtain the average pooled feature map F. avg and max pooling feature map F max Subsequently, the two feature maps are concatenated along the channel dimension ⊕, and then subjected to feature interaction and dimensionality reduction through a 1×1 convolution kernel. After nonlinear mapping using the Sigmoid activation function σ, a spatial attention weight map Δ is generated for locating weak feature regions of spores. SA Finally, the channel attention feature map Δ CA Spatial attention weight map Δ SA Perform element-wise multiplication (represented by the symbol ⊗) to obtain the final output dual attention feature map F. out .

[0094] Secondly, RFCBAMConv simulates the receptive field mechanism of the human visual system, introducing adaptive adjustment capability of the receptive field in spatial attention through multi-scale grouped convolution. This mechanism addresses the weak feature deformation problem caused by varying spore sizes, unlike the fixed square receptive field of traditional convolution. RFCBAMConv dynamically adjusts the feature response intensity at different spatial locations through grouped convolution, which is equivalent to applying a non-uniform receptive field weight distribution in different regions. This allows the network to flexibly focus according to the spore size and morphology, solving the problem of receptive field rigidity caused by parameter sharing in standard convolution, thereby adapting to the weak feature morphological changes of spores of different sizes.

[0095] The improved RFCBAMConv_C3k2 module retains residual connections through a dual-path design, ensuring that high-resolution features in the shallow layers (such as spore edge contours) are effectively transferred to the deep layers. This alleviates feature dilution of small targets after multiple downsamplings and prevents weak features from disappearing completely in the deep network. Meanwhile, experiments show that this module is almost identical to the original C3k2 in terms of parameter quantity and computational cost, but effectively improves the detection accuracy and recall rate of small targets.

[0096] In the backbone and neck networks, the original C3k2 modules in layers 2, 4, 6, 8, 13, 16, 19, and 22 were all replaced with RFCBAMConv_C3k2, forming a hierarchical weak feature enhancement mechanism. The shallow modules (P2, P3) focus on strengthening the spore edge contour through spatial attention, initially locking candidate regions in complex backgrounds, and preserving the spatial structure of weak features. The deep modules (P4, P5) use channel attention and receptive field adaptation to select high-confidence features at the semantic level, enhance the semantic discriminative power of weak features, and suppress false detections caused by impurities or uneven illumination. As a result, the multi-scale feature maps output by the backbone network have stronger discriminative power and noise resistance, and improve the expressive ability of weak feature targets.

[0097] 2.3. Enhancing weak feature recognition capabilities in complex backgrounds using a dual-branch attention module:

[0098] In the neck feature fusion network of YOLO11n, the original PAN-FPN structure achieves multi-scale feature aggregation through upsampling and channel splicing; however, for the microscopic images of silkworm microspores, the target size is extremely small (only 2~4μm in diameter), the boundary is blurred and it is highly similar to the background impurities. Such weak features face special difficulties in cross-layer fusion.

[0099] Traditional fusion strategies have three inherent limitations. First, while shallow features retain spore edge details, they lack semantic guidance and are susceptible to uneven lighting and noise interference, leading to confusion between foreground and background in the fused feature map, and the details of weak features are easily covered by noise. Second, although deep features have strong semantic discriminative power, multiple downsampling results in severe loss of spatial location information of microspores, and weak features are almost invisible in deep feature maps. Third, simple channel splicing or element-wise addition operations lack adaptive trade-offs for the importance of features at different levels. All pixels and channels are treated equally, failing to highlight the area where the spore is located, resulting in the response of weak feature areas being diluted by the background. These limitations lead to large localization errors in the multi-scale feature map output from the neck near the spore boundary, directly affecting the regression accuracy of the detection head.

[0100] To address the issues of detail loss and semantic guidance deficiency in cross-layer fusion of weak features, the RCL-YOLO model proposed in this application introduces a context-aware fusion module with an attention mechanism—CAFMFusion (Convolution and Attention Fusion Module with Pixel Attention)—into the neck network. This module is integrated into the cross-layer feature fusion process between the backbone network and the detector head. Through the dual-branch collaboration of convolution and multi-head self-attention, as well as pixel attention weighting, it achieves more refined multi-scale feature fusion, specifically designed to preserve and enhance the detailed information of weak features.

[0101] The design of the CAFMFusion module is divided into two steps;

[0102] The first step is the convolution and attention fusion module (CAFM). First, shallow features and deep features are added element-wise to obtain initial fused features, which are then fed into two branches in parallel. (1) Local branch: 1×1 convolution and Channel Shuffle operation are applied sequentially to enhance channel interaction, and then 3×3 convolution is used to capture local spatial correlation. This branch is sensitive to details such as edges and corners of weak spore features. (2) Global branch: After dimensionality reduction by 1×1 convolution, a multi-head self-attention mechanism is used to generate queries, keys and values. Values ​​are weighted and aggregated through attention maps to capture long-range dependencies across spatial locations. This branch provides semantic context for weak features and suppresses false activation of background regions. The outputs of the two branches are connected and integrated by convolution to achieve initial multi-scale, cross-level feature fusion.

[0103] The second step is pixel attention weighting. The enhanced features from the CAFM module are connected along the channel dimension to the initial fused features, and then input into a grouped convolution (7×7 convolution kernel, the number of groups equals the number of channels). A pixel-level weight map is generated by the Sigmoid activation function. This weight map is multiplied by the input shallow and deep features respectively, and spatial importance coefficients are adaptively assigned to both to complete fine-grained spatial information fusion. For example, near the edge of the spore, the weight map assigns higher coefficients to the shallow features to preserve the weak feature of fine boundary; in the uniform region inside the spore, higher coefficients are assigned to the deep features to enhance semantic consistency. Finally, a high-quality multi-scale fused feature is output through a 1×1 convolution.

[0104] In the RCL-YOLO neck structure, the CAFMFusion module is deployed at key cross-layer connections between the backbone and the head, replacing the original simple splicing operation. The multi-scale feature map optimized by this module effectively balances the preservation of spore edge details and semantic discrimination enhancement, ensuring that weak features are preserved rather than diluted during cross-layer transmission, providing more discriminative feature representations for subsequent detection heads. The structure of CAFMFusion is as follows: Figure 6 As shown.

[0105] 2.4. Improve the IOU threshold by integrating the positioning quality estimation module:

[0106] In the task of detecting silkworm microspores, the original YOLO11 detection head adopts a decoupled structure, with the classification and localization branches being independent of each other, lacking an explicit evaluation mechanism for the localization quality of bounding boxes. For spore microscopic images, the target size is extremely small and the boundaries are blurred. These weak features lead to significant uncertainty in the localization of predicted boxes, and the localization accuracy has a significant impact on the final detection result. If a localization quality estimation branch can be introduced into the detection head to quantitatively evaluate the reliability of predicted boxes and feed it back to the classification score, the detection accuracy at high IOU thresholds will be effectively improved, which is especially critical for targets with weak features and blurred boundaries.

[0107] To address the localization uncertainty caused by the blurred boundaries of weak-feature spores, the RCL-YOLO model proposed in this application introduces a localization quality estimation module—LQEHead (Localization Quality Estimation Head)—into the detection head. This module draws on the core idea of ​​the distribution-guided quality predictor in GFLV2, using bounding box distribution statistics to predict the localization quality score and perform additive calibration on the classification score. The GFL series methods introduce a general distribution to flexibly model the bounding box and design a Quality Focal Loss to jointly represent the classification score and IOU, laying the foundation for subsequent localization quality estimation. For weak-feature spores, due to the smooth transition of boundary gray levels, the distribution of the regression branch output is often relatively flat. By extracting the statistical features of this distribution, the localization quality can be effectively estimated, thereby calibrating the classification score. Figure 7 The complete structure of LQEHead is shown, including two interconnected sub-graphs, which respectively represent the evolution from the original detection head to the improved detection head and the internal details of the core module.

[0108] Figure 7 Subgraph ①Detect_LQE-Branch in this application is the improved detection head structure proposed in this application;

[0109] Based on the original structure, the output of the classification branch is fused with the output of a newly added LQE module. The input of the LQE module comes from the features of the bounding box regression branch. This design allows the classification confidence to be dynamically calibrated according to the localization quality. Figure 7 Subgraph ② LQE is the core module of LQEHead, showing its two key steps: "statistical feature extraction" and "quality estimation (MLP)". It receives the distribution vector of predicted corners (pred_corners) from the bounding box regression branch in subgraph ①, extracts features reflecting localization uncertainty through TopK statistics and Mean pooling, and maps them to the final localization quality adjustment value (quality_score) through the MLP network.

[0110] The specific calculation process for LQEHead is as follows: Figure 7 As shown in (②), its calculation process consists of three steps;

[0111] The first step is Top-K statistical feature extraction. For the four-sided (left, right, top, bottom) distribution vectors output by the bounding box regression branch, the top four maximum values ​​and their mean are extracted for each side, resulting in a 5-dimensional statistical feature for each side, for a total of 20 dimensions. The reason for choosing the Top-4 values ​​and their mean is that the sum of the distribution vectors is fixed at 1, and the Top-4 values ​​and mean can effectively reflect the flatness of the distribution—the larger the value, the sharper the distribution, and the higher the corresponding localization quality. For weak feature spores, the Top-4 value is smaller when the distribution is flat, and the quality adjustment value is correspondingly lower, thus effectively reducing the interference of low-quality predicted boxes. At the same time, this statistical feature is insensitive to the target scale and can generate a robust representation that is not affected by the size of the spores.

[0112] The second step is Softmax normalization, which involves performing Softmax normalization on the extracted statistical features along the channel dimension to convert the feature values ​​into a probability distribution form, thereby enhancing the consistency of feature representation.

[0113] The third step is MLP mapping, which feeds the normalized 20-dimensional statistical features into a lightweight MLP sub-network. The features are transformed through two pointwise convolutional layers (1×1 convolutions). The first convolutional layer maps the number of channels from 20 dimensions to the dimensions of the intermediate hidden layer and performs a non-linear transformation through the ReLU activation function. The second convolutional layer maps the output of the hidden layer to a scalar and generates the localization quality adjustment value quality_score.

[0114] Finally, LQEHead employs an additive calibration mechanism, adding the localization quality adjustment value (quality_score) to the initial classification score output by the classification branch to obtain the calibrated final classification score. This joint representation is supervised by the quality focus loss function during the training phase and directly serves as the NMS ranking criterion during the inference phase. This effectively suppresses low-quality predicted boxes caused by spore boundary blurring (one of the core manifestations of weak features) and improves detection accuracy under high IOU thresholds.

[0115] The LQEHead module has high computational efficiency, and the additional parameters introduced account for only about 0.003% of the total number of parameters in the detection model, which has almost no impact on the training and inference speed of the detector. In the RCL-YOLO ablation experiment, this module improved mAP50 by 0.43 percentage points and mAP50:95 by 0.24 percentage points without increasing the model memory usage, verifying the positive effect of the localization quality estimation mechanism on the detection of weak spore features.

[0116] 3. Experiment and Analysis:

[0117] 3.1 Experimental Environment and Evaluation Indicators:

[0118] 3.1.1 Experimental Environment and Training Configuration:

[0119] To adapt to small object detection tasks, all models use the following unified training configuration;

[0120] The input image size was fixed at 1280×1280 pixels to preserve details of small targets. The optimizer used stochastic gradient descent (SGD) with an initial learning rate of 0.01, which was decayed to 0.01 using cosine annealing. Due to memory constraints at high resolutions, the batch size was set to 8. For data augmentation, mosaic enhancement was used, but disabled in the last 15 rounds to promote stable convergence. The weight decay coefficient was set to 0.0005, and label smoothing (label_smoothing=0.0) was not used. Data caching (cache=True) was enabled during training to accelerate data loading. Eight threads were used for data loading, and all experiments were performed on a device with a single NVIDIA RTX 4090 (24GB VRAM) GPU.

[0121] 3.1.2 Evaluation Indicators:

[0122] This application uses standard evaluation metrics for object detection to evaluate the model performance;

[0123] The experimental results were quantified for precision (P), recall (R), mean precision (mAP50) with an IOU threshold of 0.5, and mean precision (mAP50:95) with an IOU threshold ranging from 0.5 to 0.95. The computational efficiency of the model was characterized by the analysis of model memory usage. This comprehensive evaluation scheme follows the general benchmarks of computer vision research to ensure the comparability of results and the standardization of methodology.

[0124] 3.2 Ablation Experiment:

[0125] To verify the effectiveness of the improved modules proposed in this application for the detection of silkworm microspores, this section uses the controlled variable method to conduct ablation experiments under the same training parameters as in Section 3.1, and quantitatively analyzes the independent contributions and synergistic gains of the three modules RFCBAMConv_C3k2, CAFMFusion, and LQEHead; the experimental results are shown in Table 1.

[0126] Table 1. Ablation experimental results for different modules:

[0127]

[0128] Experiment 1 served as the benchmark test group, using the original YOLO11n model for detection. The precision was 75.70%, the recall was 83.42%, the mAP50 was 82.44%, the mAP50:95 was 31.98%, the model memory usage was 5.4MB, and the number of parameters was approximately 2.58 million, providing a reference benchmark for subsequent improvements.

[0129] Experiment 2 only introduced the RFCBAMConv_C3k2 module, and the model memory usage increased to 5.7MB;

[0130] This module enhances the ability to focus on the edge contours of microspores and improves noise resistance in complex backgrounds by working in conjunction with a receptive field adaptive mechanism and channel spatial dual attention. Specifically, the receptive field adaptive mechanism enables the convolution kernel to dynamically adjust the sampling range according to the spore scale and morphology, and applies a non-uniform receptive field weight distribution at different spatial locations, solving the problem of receptive field rigidity caused by parameter sharing in standard convolution. Channel attention extracts multi-receptive field features through grouped convolution and generates channel weights through a spatial interaction matrix, suppressing redundant responses caused by uneven illumination and impurities. Spatial attention further compresses along the channel dimension to generate a spatial weight map, accurately identifying the high-response region where the spore is located.

[0131] After the improvements, the precision increased by 1.78 percentage points, the recall increased by 1.24 percentage points, the mAP50 increased by 0.74 percentage points, and the mAP50:95 increased by 0.36 percentage points. The experimental results show that the improved feature extraction capability of the backbone network has a direct promoting effect on the detection of weak spore features.

[0132] Experiment 3 used only the CAFMFusion module, and the model memory usage increased to 7.2MB;

[0133] This module achieves a refined fusion of shallow spatial details and deep semantic information through the dual-branch collaboration of local convolution and global self-attention, as well as pixel attention weighting. Specifically, the local branch uses 1×1 convolution, ChannelShuffle, and 3×3 convolution to enhance channel interaction and capture local spatial correlations, preserving the edge details of the spores. The global branch captures long-range dependencies across spatial locations through a multi-head self-attention mechanism, providing semantic guidance for small targets. The pixel attention module generates a spatial weight map using CAFM-enhanced features as a priori, adaptively assigning importance coefficients to shallow and deep features, thereby suppressing background noise responses.

[0134] After the improvements, the precision increased by 2.51 percentage points, mAP50 increased by 0.81 percentage points, mAP50:95 increased by 0.55 percentage points, and the recall decreased slightly by 0.16 percentage points. This indicates that after the neck feature fusion optimization, the model's regression of the bounding box is more rigorous, and some low-confidence targets are filtered out. The recall compensation needs to be achieved in conjunction with other modules.

[0135] Experiment 4 only introduced the LQEHead module, and the model memory usage was the same as the benchmark at 5.4MB;

[0136] This module uses the discrete probability distribution learned by the bounding box regression branch to extract the Top-4 values ​​and their mean as statistical features. After Softmax normalization, these features are fed into a lightweight MLP sub-network to generate localization quality adjustment values ​​and perform additive calibration on the classification scores.

[0137] After the improvement, the precision increased by 0.62 percentage points, the recall increased by 0.42 percentage points, the mAP50 increased by 0.43 percentage points, and the mAP50:95 increased by 0.24 percentage points, which verifies that the localization quality estimation mechanism has a positive effect on suppressing low-quality prediction boxes caused by spore boundary ambiguity; this module introduces only about 0.003% of additional parameters and hardly increases the computational overhead.

[0138] Further experiments with dual-module combinations verified the gain complementarity between the modules;

[0139] Experiment 5 (RFCBAMConv_C3k2+CAFMFusion) combines backbone attention enhancement with neck context awareness fusion, with a model memory footprint of 7.5MB;

[0140] The backbone module provides cleaner and more discriminative underlying features, making the complementarity of shallow details and deep semantics more efficient during neck fusion. Precision is improved by 3.19 percentage points, recall by 0.56 percentage points, mAP50 by 0.98 percentage points, and mAP50:95 by 0.72 percentage points compared to the baseline.

[0141] The memory usage of the Experiment 6 (CAFMFusion+LQEHead) model was 7.2MB;

[0142] The CAFMFusion-optimized multi-scale features provide LQEHead with a more accurate bounding box distribution, making its localization quality estimation more reliable. The accuracy is improved by 3.17 percentage points, mAP50 by 1.01 percentage points, and mAP50:95 by 0.64 percentage points.

[0143] The memory usage of the Experiment 7 (RFCBAMConv_C3k2+LQEHead) model is only 5.8MB;

[0144] Significant performance gains were achieved while maintaining a lightweight design. The efficient synergy of this combination stems from the deep functional complementarity of the two modules. RFCBAMConv_C3k2 enhances the edge contour features of spores at the underlying level through receptive field adaptation and dual attention mechanisms, making the probability distribution learned by the bounding box regression branch sharper and more reliable. LQEHead precisely utilizes the Top-4 statistics of this high-quality distribution to accurately evaluate the localization quality and effectively improves the detection accuracy under high IOU thresholds through additive calibration. The higher the quality of the backbone features, the more accurate the calibration of LQEHead. The two promote each other, resulting in an improvement of precision by 3.42 percentage points, recall by 1.84 percentage points, mAP50 by 2.35 percentage points, and mAP50:95 by 0.48 percentage points. The improvement in precision and mAP50 is the best among all dual-module combinations, fully demonstrating the high synergistic effect between backbone feature enhancement and detection head quality calibration.

[0145] Experiment 8 integrates all three modules, with a model memory usage of 7.5MB;

[0146] Optimal detection performance was achieved within acceptable limits; precision reached 79.49%, an improvement of 3.79 percentage points over the benchmark; recall reached 85.93%, an improvement of 2.51 percentage points; mAP50 reached 85.38%, an improvement of 2.94 percentage points; mAP50:95 reached 32.96%, an improvement of 0.98 percentage points. In terms of functional division, RFCBAMConv_C3k2 enhances the expression of low-level features and noise resistance in the feature extraction stage, providing high-quality input features for subsequent modules; CAFMFusion strengthens cross-layer semantic association and detail preservation in the feature fusion stage, balancing shallow spatial information and deep semantic discrimination; LQEHead improves the efficiency of bounding box quality screening in the prediction stage and suppresses the interference of low-quality predicted boxes.

[0147] The three modules complement and collaborate in the progressive detection chain of "extraction-fusion-screening," jointly achieving accurate detection of weak features of silkworm microspores. In addition, the improvement in mAP50:95 (0.98 percentage points) is greater than the contribution of any single module (0.36, 0.55, and 0.24 percentage points, respectively), indicating that the multi-module combination not only improves the detection accuracy at an IOU threshold of 0.5, but also effectively enhances the model's localization robustness at a more stringent high IOU threshold.

[0148] Referring to the visualization of the ablation experiment results shown in Figure 8, the best results were achieved when the three modules A+B+C on the far right worked together. Here, A represents RFCBAMConv_C3k2, B represents CAFMFusion, and C represents LQEHead.

[0149] 3.3 Comparative Experiment:

[0150] To further verify the detection performance of the proposed RCL-YOLO model, the proposed RCL-YOLO model is compared with current mainstream object detection models. The detection results of each model on the test set are shown in Table 2. Figure 9 P, R, and map50 curves for different detection models.

[0151] Table 2. Accuracy comparison of different models on the test set:

[0152]

[0153] To quantitatively compare model performance, the experiment used the same dataset and training parameter settings for 100 rounds of iterative training and testing, comparing the detection performance of the RCL-YOLO model of this application with that of YOLOv3n, YOLOv6n, YOLOv8n, YOLOv12n, YOLO26n, RTDETR-r18, and RTDETR-r34 models;

[0154] The experimental results are shown in Table 2. The mAP50 and mAP50:95 of the RCL-YOLO model in this application reached 85.38% and 32.96%, respectively. Compared with other algorithm models, the mAP50, mAP50:95, precision and recall are the best.

[0155] Compared to YOLOv3n, YOLOv6n, YOLOv8n, YOLOv12n, and YOLO26n models, the mAP50 and mAP50:95 improved by 2.11 and 0.78 percentage points, 4.02 and 2.28 percentage points, 3.95 and 2.33 percentage points, 11.80 and 6.27 percentage points, and 4.12 and 1.16 percentage points, respectively; precision and recall improved by 0.56 and 3.47 percentage points, 2.88 and 8.42 percentage points, 3.26 and 6.50 percentage points, 11.46 and 15.28 percentage points, and 6.20 and 8.26 percentage points, respectively; compared to the Transformer architecture detection model RTD... The mAP50 and mAP50:95 of ETR-r18 and RTDETR-r34 were improved by 14.44 and 10.94 percentage points, and 7.45 and 7.87 percentage points, respectively, while the precision and recall were improved by 6.39 and 10.17 percentage points, and 3.68 and 6.19 percentage points, respectively. The improvement in mAP indicates that the improved model can more accurately integrate shallow details such as spore edge texture with deep semantic features such as the overall morphology of spores, enhancing the multi-scale representation ability of extremely small targets. The improvement in precision indicates that the improved model has a lower false detection rate and can effectively suppress background noise interference such as impurities and bubbles in microscopic images. The improvement in recall indicates that the improved model has significantly reduced the missed detection of low-contrast spores.

[0156] Furthermore, the RCL-YOLO model has a memory footprint of only 7.5MB, which is significantly lower than lightweight models such as YOLOv8n and YOLOv12n, as well as Transformer architecture models such as RTDETR-r18 and RTDETR-r34. At the same time, the detection speed reaches 57.71f / s, which meets the requirements of real-time detection. This indicates that the model in this application has achieved a good balance between detection accuracy, model size and inference speed, and is suitable for deployment in grassroots silkworm seed quality inspection scenarios.

[0157] To further compare RCL-YOLO with other detection models, evaluations were conducted on the validation set, and the results are shown in Table 3. It can be seen that RCL-YOLO outperforms other detection models on the validation set for detecting silkworm microspores, achieving an mAP50 of 87.09% and a detection speed of 61.40 f / s. YOLOv3n demonstrates better detection accuracy than YOLOv6n, YOLOv8n, YOLOv12n, YOLO26n, RTDETR-r18, RTDETR-r34, and Faster RCNN, with an mAP50 of 83.66%, but its detection speed is only 31.95 f / s, 29.45 f / s lower than RCL-YOLO. RTDETR-r18 and RTDETR-r34 have mAP50:95 of 23.78% and 24.88%, respectively, and compared to the RCL-YOLO model, they are not suitable for the high-precision requirements of silkworm microspore detection.

[0158] In addition, through Figure 9 It can be seen that RCL-YOLO outperforms other detection models in both precision (P) and recall (R). The mAP50 curve rises rapidly in the first 40 epochs and then stabilizes after 70 epochs, still outperforming other detection models. This indicates that the improved model also has good convergence efficiency and detection stability on the validation set.

[0159] Table 3. Accuracy comparison of different models on the validation set:

[0160]

[0161] 3.4 Accuracy Analysis of Silkworm Microspore Detection at Different Image Resolutions:

[0162] The detection performance of the model is affected by both image resolution and data volume. This section explores the effect of different resolutions and different data volumes on detection accuracy, and verifies the robustness and generalization ability of the RCL-YOLO model.

[0163] From a research perspective, the choice of resolution directly affects the model's ability to perceive weak features of spores. Too low a resolution will lead to the loss of spore edge details and insufficient feature expression, especially with a significant decrease in localization accuracy at high IOU thresholds. While too high a resolution can retain more spatial details, the proportion of spore pixels will decrease accordingly, and background impurities and noise information such as bubbles will increase simultaneously, which may interfere with the model's learning of weak features.

[0164] From a production application perspective, grassroots silkworm seed quality inspection scenarios require the system to balance detection accuracy and real-time performance under conventional hardware conditions, as summarized in Table 4 and... Figure 10The experimental results show that 1280×1280 pixels represents the optimal balance between detection performance and computational efficiency under the conditions of this study. At this resolution, as the data volume increases from 361 to 1275 images, mAP50 increases from 85.10% to 85.71%, and mAP50:95 increases from 32.23% to 32.85%, with a detection speed of 57.71 f / s, which can meet the real-time detection needs of grassroots units. In contrast, although 1000×1000 pixels have a lower computational load, the mAP50:95 decreases significantly (only 31.26% with 1235 images), indicating that the positioning accuracy is limited under high IOU thresholds. Although 1500×1500 pixels have higher accuracy with the same data volume, the accuracy decreases with increasing data volume (mAP50 decreases from 86.54% to 86.12% when the data volume increases from 320 to 1150 images), and the computational cost increases, requiring higher hardware deployment.

[0165] Therefore, this study recommends 1280×1280 pixels as the optimal input resolution for the silkworm microspore detection task, which ensures effective expression of weak features while also meeting the real-time requirements in grassroots scenarios. This study used 1280×1280 pixel images and the original acquired images for subsequent experiments, and then collected new microscopic spore images with different resolutions and data volumes for comparative testing. The results are shown in Table 4, and the detection results are visualized as follows: Figure 10 As shown.

[0166] Table 4. Detection results for multiple datasets at different resolutions:

[0167]

[0168] Through Table 4 and Figure 10 It can be observed that at resolutions of 1000×1000 pixels and 1280×1280 pixels, the model's mAP50 and mAP50:95 are further improved with the increase of data volume;

[0169] In Figure 11, at a resolution of 1000×1000 pixels, when the number of data images increased from 318 to 1235, the mAP50 gradually increased from 83.10% to 84.21%, an increase of 1.11 percentage points; at a resolution of 1280×1280 pixels, when the number of data images increased from 361 to 1275, the mAP50 increased from 85.10% to 85.71%, an increase of 0.61 percentage points. This indicates that at lower resolutions, increasing the amount of training data can effectively improve the detection accuracy of the model. However, at a resolution of 1500×1500 pixels, as the number of images increases, both mAP50 and mAP50:95 show an overall trend of decreasing accuracy with increasing data volume.

[0170] Figure 11 visually illustrates that at a resolution of 1500×1500 pixels, the mAP50 is 86.54% when the number of data images is 320, drops to 86.28% when the number of data images increases to 550, and further decreases to 86.12% when the number of data images increases to 1150. This is because the overall target of silkworm microspores is smaller at a resolution of 1500×1500 pixels, and with the increase of the amount of microscopic image data, noise such as impurities and bubbles also increases. Therefore, at a resolution of 1500×1500 pixels, the mAP50 and mAP50:95 show a certain downward trend with the increase of the number of images.

[0171] Meanwhile, comparing the detection accuracy at the three resolutions, Figure 11 clearly shows that the mAP50 curve at the 1500×1500 pixel resolution is generally above that at the 1000×1000 and 1280×1280 pixel resolutions. The detection accuracy at the 1500×1500 pixel resolution is generally better than the other two. This is because the spore target features to be detected are weakened to some extent at the 1000×1000 and 1280×1280 pixel resolutions. On the other hand, the lower resolution input causes the target to be detected to become larger and smaller in scale, resulting in a mismatch between the detection head and the target distribution. Therefore, the detection accuracy of the other two resolutions is slightly lower than that of the 1500×1500 pixel resolution. At the 1000×1000 pixel resolution, the decrease in mAP50:95 is more significant than that of mAP50. This is because the proportion of spore edge pixels is smaller at lower resolutions, and the regression accuracy of the prediction box is more sensitive to the IOU threshold. Even a small offset of the bounding box will cause the matching to fail at a high IOU threshold, resulting in a lower mAP50:95.

[0172] Furthermore, comparing the detection accuracy at the same data volume under three different resolutions, from Figure 10 As can be seen, the overall difference in detection accuracy between 1000×1000 pixels and 1280×1280 pixels is relatively small, while the detection accuracy at 1500×1500 pixels is higher. Taking a medium amount of data (about 550~650 images) as an example, the mAP50 is 83.59% at 1000×1000 pixels, 85.43% at 1280×1280 pixels, and 86.28% at 1500×1500 pixels. The higher resolution improves by 2.69 percentage points compared to the lower resolution.

[0173] Figure 10 The vertical axis is divided into three levels from left to right according to the amount of data:

[0174] Small (approximately 300-360 pieces), Medium (approximately 550-660 pieces), Large (approximately 1150-1275 pieces);

[0175] The figures clearly show that at 1000×1000 and 1280×1280 resolutions, mAP50 continuously increases with increasing data volume, with increases of 1.11 and 0.61 percentage points respectively, indicating that increasing training samples is beneficial for improving detection accuracy at low resolutions. However, at a high resolution of 1500×1500, mAP50 decreases with increasing data volume (from 86.54% to 86.12%), which is related to the fact that noise (impurities, bubbles) in high-resolution images increases synchronously with the amount of data, making it difficult for the model to distinguish weak features from noise. Furthermore, with the same amount of data, the mAP50 at 1500×1500 resolution is consistently higher than the other two resolutions, indicating that appropriately increasing the resolution helps preserve the edge details of spores, but attention should be paid to controlling noise interference. The above analysis shows that there is a trade-off between resolution and data volume, and in practical applications, an appropriate input resolution should be selected based on the quality of the dataset.

[0176] 3.5. Silkworm Microspore Detection System:

[0177] To facilitate intelligent detection and quantitative statistics of silkworm microspores, an intelligent detection system for silkworm microspores based on the RCL-YOLO model proposed in this application was developed using the Flask framework and a web front-end. The system includes five functional modules: model management, parameter adjustment, file upload, real-time monitoring, and statistical output. Users can load pre-trained weights through a webpage and adjust the confidence level and IOU threshold in real time. It supports three input modes: single image, offline video, and real-time microscope footage (acquired via an OBS virtual camera). Detection results are displayed in real time with blue bounding boxes and confidence level labels, and statistical information such as spore count, processing time, and average confidence level are automatically summarized.

[0178] As shown in Figure 11(a), the system detects a single microscopic image, displaying the original image on the left and the detection results in real time on the right. A total of 25 spore targets were detected, taking approximately 330ms. Figure 11(b) shows the offline video detection function, where the system infers the video frame by frame and detects a total of 646 spore targets. Figure 11(c) shows the real-time monitoring mode, which pushes the real-time image of the microscope through an MJPEG stream to achieve frame-by-frame detection with a frame rate of 20~30FPS.

[0179] Experimental verification shows that the system has good real-time performance and accuracy in image, video and real-time monitoring modes, meeting the rapid detection needs of grassroots silkworm seed quality inspection scenarios. The system adopts a browser / server architecture, which does not require the installation of client software, lowering the threshold for using deep learning models and providing a convenient technical tool for intelligent silkworm seed quality inspection.

[0180] 4. Discussion:

[0181] This application addresses the challenge of weak feature detection in microscopic images of silkworm microspores, which are characterized by extremely small target size, blurred boundaries, and high similarity to background impurities. It proposes an RCL-YOLO model based on an improved YOLOv11n. Experimental results show that this model outperforms several mainstream detection algorithms on a self-built dataset. This section discusses the model improvement mechanism, experimental results, and potential problems.

[0182] 4.1 Effectiveness and Collaboration Mechanism of Module Improvement:

[0183] Ablation experiments (Table 1) show that all three improved modules contribute positively to the detection performance;

[0184] Among them, the RFCBAMConv_C3k2 module effectively solves the problem of fixed receptive fields and difficulty in focusing on the edges of microspores in standard convolution by using an adaptive receptive field mechanism and channel-space dual attention. In Experiment 2, introducing this module alone improved mAP50 by 0.74 percentage points, verifying its effect on improving the signal-to-noise ratio of weak features. The CAFMFusion module alleviates the conflict between shallow details and deep semantics in cross-layer fusion by using the collaboration of local and global branches and pixel attention weighting. In Experiment 3, using this module alone improved mAP50 by 0.81 percentage points, but the recall rate decreased slightly (-0.16 percentage points). This may be because the fused features have stricter requirements for bounding box regression, and some low-confidence spores are filtered out. The LQEHead module uses the statistical features of bounding box distribution to estimate the localization quality and performs additive calibration on the classification score, improving mAP50 by 0.43 percentage points without increasing the number of parameters, especially with a significant advantage at high IOU thresholds.

[0185] It is worth noting that the RFCBAMConv_C3k2+LQEHead (Experiment 7) dual-module combination achieved a 2.35 percentage point improvement in mAP50 with only 5.8MB of memory usage, indicating a high degree of synergy between backbone feature enhancement and head quality calibration. The clearer underlying features also provide a more reliable distribution basis for localization quality estimation.

[0186] 4.2 Discussion on resolution, data volume, and system deployment:

[0187] The experiments in Section 3.4 show that increasing the amount of training data can steadily improve detection accuracy at lower resolutions (1000×1000, 1280×1280), while at a higher resolution of 1500×1500, the accuracy decreases with increasing data volume.

[0188] This phenomenon can be analyzed from two aspects:

[0189] On the one hand, high resolution can preserve more details of spore edges. Therefore, with the same amount of data, the detection accuracy of 1500×1500 resolution (mAP50 of 86.28%~86.54%) is consistently better than 1000×1000 (83.59%~84.21%) and 1280×1280 (85.10%~85.71%). However, high resolution also amplifies noise signals such as impurities and bubbles in the background. When the amount of data increases, the growth rate of noise diversity exceeds the model's ability to learn weak features, causing the accuracy to decrease with the increase of data volume (from 320 images to 115 images). At 0 frames per second, mAP50 decreased from 86.54% to 86.12%; on the other hand, at 1000×1000 resolution, the decrease in mAP50:95 was more significant than that of mAP50, indicating that lower resolution has a greater impact on positioning accuracy at high IOU thresholds. This is because the proportion of spore edge pixels is smaller at lower resolutions, and the bounding box regression accuracy is more sensitive to the IOU threshold. In summary, there is a trade-off between resolution and data volume. 1280×1280 pixels achieves the optimal balance between detection accuracy, data utilization efficiency, and computational overhead, making it the recommended configuration for grassroots silkworm seed quality inspection scenarios.

[0190] Based on the above analysis, this study further developed an intelligent detection system based on Flask and a web front-end (Section 3.5). This system performs stably in image, video, and real-time monitoring modes, with a single image processing time of approximately 330ms and a real-time monitoring frame rate of 20-30 FPS. The system adopts a browser / server architecture, eliminating the need for client software installation and lowering the barrier to entry for grassroots quality inspectors. The RCL-YOLO model has a memory footprint of only 7.5MB, significantly lower than comparable models such as YOLOv3n (792MB) and RTDETR-r34 (119.9MB), allowing deployment on low-configuration office computers and demonstrating good hardware compatibility. The real-time monitoring module acquires microscope images via an OBS virtual camera, enabling frame-by-frame inference and MJPEG stream delivery, meeting the real-time observation needs of grassroots personnel. Experiments show that the system maintains stable detection performance for both sparsely and densely distributed spore samples, effectively reducing the labor intensity of manual microscopic examination and improving the objectivity of the detection results.

[0191] 5. Conclusion:

[0192] This application addresses the challenge of weak feature detection in microscopic images of silkworm microspores, which exhibit extremely small target size, blurred boundaries, and high similarity to background impurities. It proposes a spore detection model based on the improved YOLOv11n—RCL-YOLO. The main conclusions are as follows:

[0193] RCL-YOLO has achieved an effective improvement in detection accuracy through three synergistic improvements;

[0194] The RCL-YOLO model incorporates the RFCBAMConv_C3k2 module into the backbone network to enhance its ability to capture weak spore features. The CAFMFusion module is introduced into the neck network to achieve refined fusion of shallow details and deep semantics. The LQEHead module is introduced into the detection head to suppress low-quality predicted boxes through localization quality calibration. On a self-built dataset of 3200 microscopic images, RCL-YOLO achieves mAP50 and mAP50:95 of 85.38% and 32.96%, respectively, representing improvements of 2.94 and 0.98 percentage points compared to the original YOLOv11n. Precision and recall are improved by 3.79 and 2.51 percentage points, respectively, with a model memory footprint of only 7.5MB, meeting the real-time detection needs of grassroots silkworm seed quality inspection scenarios.

[0195] Ablation experiments verified the independent contributions and synergistic gains of the three modules;

[0196] In single-module experiments, RFCBAMConv_C3k2, CAFMFusion, and LQEHead improved mAP50 by 0.74, 0.81, and 0.43 percentage points, respectively. In dual-module combinations, the combination of RFCBAMConv_C3k2 and LQEHead achieved a 2.35 percentage point improvement in mAP50 at the cost of only 0.4 MB, which was the best among all dual-module combinations. After the three-module integration, mAP50:95 improved by 0.98 percentage points, indicating that the multi-module combination effectively improved the robustness of the model under high IOU thresholds.

[0197] The test results of multiple datasets at different image resolutions (1000×1000, 1280×1280, 1500×1500 pixels) show that RCL-YOLO exhibits strong detection stability at different resolutions and data volumes. The detection accuracy at 1500×1500 pixels is generally better than that at lower resolutions, but the accuracy decreases with the increase of data volume, indicating that attention should be paid to noise interference at high resolutions.

[0198] A smart detection system for silkworm microspores based on Flask and Web front-end was designed and developed;

[0199] The system uses RCL-YOLO as its inference engine and integrates five major functional modules: model management, parameter adjustment, file upload, real-time monitoring, and statistical output. The real-time monitoring module acquires microscope images through an OBS virtual camera and performs frame-by-frame inference. The system has good real-time performance and accuracy in image, video, and real-time monitoring modes, providing a convenient and efficient technical tool reference for the intelligent inspection of silkworm seed quality at the grassroots level.

[0200] Future research may also explore the following directions: lightweight feature fusion strategies, such as using depthwise separable convolutions to replace some self-attention operations; introducing multispectral data fusion to improve the distinguishability of spores in complex backgrounds; combining knowledge distillation and other techniques to compress the model to a smaller size while maintaining accuracy, making it compatible with more low-cost embedded hardware; and further expanding the dataset of microscopic images in multiple scenarios and categories to verify the model's generalization ability in a wider range of tasks. Further details will not be elaborated upon here.

[0201] The above embodiments can be implemented, in whole or in part, by software, hardware, firmware, or any other combination thereof. When implemented in software, the above embodiments can be implemented, in whole or in part, as a computer program product. Those skilled in the art will recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution.

[0202] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment, depending on actual needs.

[0203] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any changes or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application.

Claims

1. A rapid detection method for silkworm microspores based on AI recognition, characterized in that, The method includes: acquiring an image dataset of the target spores and performing data preprocessing; Based on the existing YOLOv11n framework, an RCL-YOLO model suitable for detecting weak spore features during microscopic examination is constructed: In the backbone network, the RFCBAMConv_C3k2 module, which integrates receptive field adaptation and channel space dual attention, is introduced to replace the original C3k2 structure. In the neck feature fusion stage, the context-aware dual-branch attention fusion module CAFMFusion is introduced to replace the original operation. Through the collaboration of local convolution and global self-attention, as well as pixel-level weighting, it achieves the fusion of shallow spatial details and deep semantic information. In the detection head, the localization quality estimation head LQEHead is introduced to explicitly evaluate the localization accuracy of the bounding box and calibrate the classification confidence.

2. The method for rapid detection of silkworm microspores based on AI recognition according to claim 1, characterized in that: The image dataset was acquired by taking pictures with an optical microscope, using the microscope and its accompanying camera, and the target spores were photographed under different lighting and density conditions.

3. The method for rapid detection of silkworm microspores based on AI recognition according to claim 1, characterized in that: When preprocessing image datasets, the enhancement strategies employed include at least: horizontal and vertical flipping, small-angle random rotation, scaling, and random brightness; where the small angle is a preset angle.

4. The method for rapid detection of silkworm microspores based on AI recognition according to claim 1, characterized in that: The RFCBAMConv_C3k2 module replaces the standard convolution within the Bottleneck in the original C3k2 with RFCBAMConv; RFCBAMConv is derived by fusing CBAM improvements on RFAConv; CBAM is a concatenated mechanism of channel attention and spatial attention.

5. The method for rapid detection of silkworm microspores based on AI recognition according to claim 4, characterized in that: In CBAM, channel attention (CA) is used to improve the signal-to-noise ratio of weak features. It extracts multiple receptive field features in parallel through grouped convolutions of different scales, and generates channel weights through spatial interaction matrix and Softmax normalization to highlight the weak spore signal in the channel dimension. Spatial attention (SA) is used to locate weak feature regions. The feature map after channel attention weighting is average pooled and max pooled along the channel dimension, and after concatenation, a spatial weight map is generated through 1×1 convolution and Sigmoid activation function to mark the high response region where the spore is located.

6. The method for rapid detection of silkworm microspores based on AI recognition according to claim 4, characterized in that: RFCBAMConv simulates the receptive field mechanism of the human visual system, introducing adaptive adjustment capability of the receptive field in spatial attention through multi-scale grouped convolution. RFCBAMConv dynamically adjusts the feature response intensity at different spatial locations through grouped convolution, which is equivalent to applying a non-uniform receptive field weight distribution in different regions. This allows the network to flexibly focus according to spore scale and morphology, adapting to the weak feature morphological changes of spores of different sizes. The improved RFCBAMConv_C3k2 module retains residual connections through a dual-path design. In the backbone and neck networks, the original C3k2 modules in layers 2, 4, 6, 8, 13, 16, 19, and 22 are replaced with RFCBAMConv_C3k2 to form a hierarchical weak feature enhancement mechanism. The shallow modules P2 and P3 enhance the spore edge contour through spatial attention, initially locking candidate regions in complex backgrounds and preserving the spatial structure of weak features. The deep modules P4 and P5 use channel attention and receptive field adaptation to select high-confidence features at the semantic level to enhance the semantic discriminative power of weak features.

7. The rapid detection method for silkworm microspores based on AI recognition according to claim 1, characterized in that: The design of the dual-branch attention fusion module CAFMFusion consists of two steps: The first step is the convolution and attention fusion module CAFM; firstly, shallow features and deep features are added element-wise to obtain initial fused features, which are then fed into two branches in parallel: In the local branch, 1×1 convolution and Channel Shuffle operation are applied sequentially to enhance channel interaction, and then 3×3 convolution is used to capture local spatial correlation. In the global branch, after dimensionality reduction using 1×1 convolution, a multi-head self-attention mechanism is used to generate queries, keys, and values. The values ​​are weighted and aggregated through an attention map to capture long-range dependencies across spatial locations. The outputs of the two branches are connected and integrated through convolution to achieve initial multi-scale, cross-level feature fusion. The second step is pixel attention weighting; The enhanced features from the CAFM module are connected to the initial fused features along the channel dimension, and then input into the grouped convolution, i.e., a 7×7 convolution kernel, with the number of groups equal to the number of channels. A pixel-level weight map is generated through the Sigmoid activation function. This weight map is multiplied by the input shallow and deep features respectively, and spatial importance coefficients are adaptively assigned to both to complete fine-grained spatial information fusion. Finally, a 1×1 convolution is performed to output multi-scale fused features; In the RCL-YOLO neck structure, the dual-branch attention fusion module CAFMFusion is deployed at the key cross-layer connection between the backbone and the head.

8. The method for rapid detection of silkworm microspores based on AI recognition according to claim 1, characterized in that: The Location Quality Estimation Head (LQEHead) predicts location quality scores by utilizing bounding box distribution statistics and performs additive calibration on classification scores, employing the core logic of the distribution-guided quality predictor in GFLV2. The complete structure of LQEHead includes: Detect_LQE-Branch, which is an improved detection head structure. Based on the original structure, it fuses the output of the classification branch with the output of a newly added LQE module. The input of the LQE module comes from the features of the bounding box regression branch. LQE is the core module of LQEHead, showing its internal statistical feature extraction and quality estimation MLP steps. It receives the predicted corner distribution vector from the bounding box regression branch in Detect_LQE-Branch, extracts features reflecting localization uncertainty through TopK statistics and Mean pooling, and maps them to the final localization quality adjustment value through the MLP network.

9. The method for rapid detection of silkworm microspores based on AI recognition according to claim 8, characterized in that: The calculation process of LQEHead consists of three steps: The first step is TopK statistical feature extraction. For the four-sided distribution vector output by the bounding box regression branch, the first four maximum values ​​and their mean are extracted for each side, resulting in 5-dimensional statistical features for each side, for a total of 20 dimensions for the four sides. The second step is Softmax normalization, which performs Softmax normalization on the extracted statistical features along the channel dimension, converting the feature values ​​into a probability distribution form. The third step is MLP mapping, which feeds the normalized 20-dimensional statistical features into a lightweight MLP sub-network and performs feature transformation through two pointwise convolutional layers. The first convolutional layer maps the number of channels from 20 dimensions to the intermediate hidden layer dimension and performs a non-linear transformation using the ReLU activation function. The second convolutional layer maps the hidden layer output to a scalar to generate a localization quality adjustment value. Finally, LQEHead uses an additive calibration mechanism to add the localization quality adjustment value to the initial classification score output by the classification branch to obtain the calibrated final classification score.

10. A rapid detection system for silkworm microspores based on AI recognition, applied to the rapid detection method for silkworm microspores based on AI recognition as described in any one of claims 1-9, characterized in that, The system includes: Data acquisition module: Acquires image datasets of target spores and performs data preprocessing; Model building module: Based on the existing YOLOv11n framework, an RCL-YOLO model suitable for detecting weak spore features during microscopic examination is constructed. In the backbone network, the RFCBAMConv_C3k2 module, which integrates receptive field adaptation and channel space dual attention, is introduced to replace the original C3k2 structure. In the neck feature fusion stage, the context-aware dual-branch attention fusion module CAFMFusion is introduced to replace the original operation. Through the collaboration of local convolution and global self-attention, as well as pixel-level weighting, it achieves the fusion of shallow spatial details and deep semantic information. In the detection head, the localization quality estimation head LQEHead is introduced to explicitly evaluate the localization accuracy of the bounding box and calibrate the classification confidence.

Images