Astragalus membranaceus decoction piece surface defect detection method based on improved YOLO system detection model

Abstract

The invention discloses an astragalus membranaceus decoction piece surface defect detection method based on an improved YOLO series model. The method comprises the following steps of: acquiring a high-definition image of the radix astragali decoction pieces through a hardware acquisition device of the radix astragali decoction pieces, constructing a labeling data set containing three defects of cracks, mildew spots and damage, and performing data enhancement through modes of rotating, overturning, cutting, brightness disturbance and the like; aiming at the characteristics of complex surface texture and irregular defect form of the astragalus membranaceus decoction pieces, a YOLO series network structure is improved, linear deformable convolution is introduced into a trunk feature extraction network, and partial convolution and gated linear unit convolution are further introduced into a feature fusion module, so that the feature extraction capability and the reasoning speed are improved; visual software constructed based on Flask and the like can display detection videos, defect statistics and defect position numbers in real time, automatic screening of the astragalus membranaceus decoction pieces in a production line is achieved, the defect detection precision and detection efficiency can be effectively improved, and the method is suitable for quality control and intelligent transformation of astragalus membranaceus decoction piece production enterprises.

Description

Technical Field

[0001] This invention relates to the field of intelligent detection technology for traditional Chinese medicine decoction pieces, specifically a method for detecting surface defects in Astragalus membranaceus decoction pieces based on an improved YOLO system detection model. Background Technology

[0002] Astragalus, a traditional Chinese medicinal herb, is widely used in clinical and health care fields. However, its processed slices are prone to surface defects such as cracks, mold, and breakage during various processing stages. These defects not only affect the appearance quality but may also reduce the efficacy of the medicine. Therefore, it is necessary to establish an accurate, stable, and rapid testing system. Currently, Chinese herbal medicine processing enterprises still rely on manual visual inspection. The efficiency of this method depends heavily on the experience of the operators and is easily affected by fatigue, oversights, and subjective differences, making it difficult to meet the standardization requirements of large-scale production.

[0003] With the development of intelligent manufacturing, the industry urgently needs automated inspection solutions to improve the reliability of quality inspection. Although traditional image processing algorithms can achieve basic automatic inspection, they still have significant limitations when faced with the complex textures and diverse defect morphologies on the surface of Astragalus membranaceus slices. The manually extracted features such as edges and grayscale are difficult to accurately characterize the defect characteristics, especially for small-area, low-contrast defects with low recognition rates. Furthermore, defects in Astragalus membranaceus slices often exhibit irregular shapes and minute scales, further increasing the difficulty of inspection.

[0004] In recent years, deep learning-based object detection technology has become mainstream in industrial inspection due to its adaptive feature extraction and high generalization capabilities, significantly improving detection performance in complex scenarios. However, existing general-purpose models still face three major technical challenges in the detection of Astragalus membranaceus (Huangqi) slices: the difficulty in capturing irregular defect features, the susceptibility of complex surface textures to false detections, and the inability to meet the real-time requirements of high-throughput production. Therefore, it is necessary to optimize the detection framework based on defect characteristics, focusing on strengthening the feature extraction module and improving the representation capability of small targets, while optimizing inference speed, to build an automated detection system suitable for the quality inspection of traditional Chinese medicine slices. Summary of the Invention

[0005] To address the limitations of traditional methods for detecting surface defects in Astragalus membranaceus slices, this invention proposes a method for detecting surface defects in Astragalus membranaceus slices based on an improved YOLO system detection model, thereby achieving high-precision, high-robustness, and high-real-time detection of surface defects in Astragalus membranaceus slices.

[0006] The technical solution to achieve the objective of this invention is: A method for detecting surface defects of Astragalus membranaceus slices based on an improved YOLO series model includes an Astragalus membranaceus slice acquisition module, a dataset construction module, an image processing module, and a result visualization module. The detection method includes the following steps: s1. Design and construct a hardware acquisition device for Astragalus membranaceus slices to acquire high-definition images of Astragalus membranaceus slices; s2. Based on high-resolution images of Astragalus membranaceus slices, construct an Astragalus membranaceus slice dataset for model training; s3. Improved YOLO series models trained based on Astragalus membranaceus decoction pieces dataset; s4. Develop visualization software based on the Flask framework and the trained detection model.

[0007] Step s1 describes designing and constructing a hardware acquisition device for Astragalus membranaceus slices to acquire high-definition images of Astragalus membranaceus slices, including: (1) Use high-resolution industrial cameras as imaging devices and configure controllable ring LED light sources to reduce ambient light interference. Adjust the brightness of the light source, the illumination angle and the camera exposure parameters to keep the acquired images high-definition and save the captured images in lossless BMP and other formats. (2) Design and use an automated conveyor belt composed of parallel square grids, each grid carrying a single piece of Astragalus membranaceus, to ensure that the Astragalus membranaceus slices are in a fixed position, with a stable spacing and no overlap during the conveying process.

[0008] Step s2, which involves constructing a dataset of Astragalus membranaceus slices for model training based on high-resolution images of Astragalus membranaceus slices, includes: (1) Perform quality screening on the original acquired images and remove image samples that are blurry, severely occluded, or abnormally exposed; (2) Perform image preprocessing on the screened images, including color correction and noise suppression, to ensure sample consistency; (3) Accurately label the Astragalus membranaceus slices using LabelImg or other annotation tools, and save the labels in a TXT file format suitable for YOLO series models; the specific dataset creation process is as follows: 1) Defects are categorized based on surface defect characteristics of Astragalus membranaceus slices. The specific discriminant characteristics are as follows: ① Crack: Fine, elongated linear fracture structures appearing on the surface of Astragalus membranaceus slices, mostly manifested as tiny cracks distributed along the grain direction or laterally; ② Mold: Abnormally colored spots or patchy moldy areas that form on the surface of Astragalus membranaceus slices, usually appearing as dark, brown or black spots; ③ Breakage: Astragalus slices show structural damage such as edge chips, local corner chipping, and obvious breakage; 2) For each image of Astragalus membranaceus slices, surface defects were annotated using the least bounding matrix method, and the defect category was noted. 3) Perform data augmentation operations on the labeled dataset using rotation, flipping, cropping, and brightness perturbation; 4) The augmented dataset is randomly partitioned into training, validation and test sets in a ratio of 7:2:1.

[0009] Step s3, which involves training an improved YOLO series model based on the Astragalus membranaceus decoction pieces dataset, includes: (1) Linear deformable convolutional modules are used to replace standard convolutional layers in the backbone feature extraction network of the model, thereby enhancing the feature capture capability for irregular defects; (2) In the feature fusion module of the YOLO series models, the PGLUBlock module is introduced. The amount of computation is reduced by introducing partial convolution PConv, and gated linear unit convolution ConvGLU is used to suppress irrelevant texture features. (3) Set the training parameters, including batch size of 32, training rounds of 300, initial learning rate of 0.01, final learning rate ratio of 0.01, optimization algorithm of SGD, and weight decay of 0.0005. Use this set of parameters to train the improved model.

[0010] Step s4 describes building visualization software based on the Flask development framework and the trained detection model, including: (1) Initialize the Flask development framework and load the weight file of the improved model training results; (2) Construct the control area, including function buttons for model selection, save path, start detection, and end detection; (3) Construct a real-time detection display area to display the video stream captured by the industrial camera in real time and overlay the defect recognition box detected by the model; (4) Construct a defect detection quantity statistics area to count the number of defects detected in real time, including cracks, mold spots and damage; (5) Construct the defect location area, map the defect coordinates output by the model to the conveyor belt grid coordinate system, and output the grid number where the defective Astragalus membranaceus slice is located.

[0011] The advantages of this invention are: 1. Fast detection speed: Compared with traditional detection methods, this invention does not require manual screening of each piece, nor does it rely on cumbersome optical special algorithm processing. It can achieve millisecond-level real-time detection in continuous production lines, with a detection time of only about 3ms for a single image.

[0012] 2. High detection accuracy: This invention optimizes and improves upon the YOLO series model framework, effectively enhancing the extraction capability of surface defect features of Astragalus membranaceus slices. The improved model's mAP50 and mAP50-95 are improved by more than 0.7% and 1.3% respectively compared to the original YOLO series model.

[0013] 3. Low actual deployment cost: This invention can be directly deployed in ordinary industrial cameras and standard light source environments without relying on expensive special optical inspection instruments; combined with the grid-like conveyor belt structure, this invention can replace manual inspection of each piece, significantly reducing labor costs.

[0014] 4. High scalability: The detection system of this invention is built on a scalable deep learning framework, allowing for flexible model replacement or upgrades, and seamless integration of the software interface with different industrial environments. This method is not only applicable to Astragalus membranaceus slices but can also be extended to the appearance quality inspection of other Chinese herbal medicine slices, demonstrating broad engineering adaptability and industrial promotion value. Attached Figure Description

[0015] Figure 1 This is a general block diagram of the detection method of the present invention; Figure 2 This is a flowchart of the detection process for Astragalus membranaceus slices according to the present invention; Figure 3 This is a diagram of the Astragalus membranaceus slice image acquisition device of the present invention; Figure 4 This is a diagram of the improved model network structure for this invention; Figure 5 This is a structural diagram of the LDConv module of the present invention; Figure 6 This is a structural diagram of the improved feature fusion module of the present invention; Figure 7 This is a design diagram of the visual software interface for this invention. Detailed Implementation

[0016] The technical solutions of the present invention will be clearly and thoroughly described below with reference to the accompanying drawings. The described embodiments are merely some embodiments of the present invention, and the specific technical solutions adopted are as follows: A method for detecting surface defects of Astragalus membranaceus slices based on an improved YOLO series model includes an Astragalus membranaceus slice acquisition module, a dataset construction module, an image processing module, and a result visualization module.

[0017] The Astragalus membranaceus (Huangqi) slice acquisition module aims to obtain high-quality raw image data. This module uses a high-resolution industrial camera as the core imaging device, paired with an automated conveyor belt to form a continuous acquisition system. The specially designed conveyor belt consists of a series of parallel square grids, each dedicated to carrying a single Astragalus membranaceus slice, maintaining a fixed position and stable spacing during transport and avoiding overlap, tilting, or jumping between slices. Furthermore, this conveyor belt structure facilitates subsequent grid-by-grid location and removal of defective slices based on inspection results, improving the automatic screening efficiency of defective slices on the production line. Simultaneously, to reduce ambient light interference, this module is equipped with a controllable ring LED light source. The brightness, illumination angle, and camera exposure parameters can be adjusted according to production line conditions to ensure image clarity. The Astragalus membranaceus slice acquisition module provides high-quality image input for subsequent modules and is a crucial foundation for the performance of the entire inspection method. The Astragalus membranaceus slice acquisition device is as follows: Figure 3 As shown.

[0018] The dataset construction module selects, preprocesses, and finely annotates the acquired raw images. First, the raw images undergo quality screening, removing unqualified samples such as those that are blurry or severely occluded. Then, the images are preprocessed, including color correction and noise suppression, to improve sample consistency. This module manually annotates Astragalus membranaceus slices using LabelImg or other annotation tools, with annotation categories including cracks, mold spots, and damage. The annotations are saved in a TXT file format suitable for YOLO series models. To enhance data diversity and improve model generalization ability, this module further employs various data augmentation strategies, including rotation, flipping, cropping, and brightness perturbation, to enhance the data and enable the model to better adapt to stable detection under complex conditions. The specific dataset creation process is as follows: Defects are categorized based on surface defect characteristics of Astragalus membranaceus slices. The specific discriminant features are as follows: (1). Crack: The fine, linear fracture structure that appears on the surface of Astragalus membranaceus slices, mostly manifested as fine cracks distributed along the texture direction or laterally.

[0019] (2). Mold: Abnormal spots or patches of mold that form on the surface of Astragalus membranaceus slices, usually dark, brown or black.

[0020] (3). Break: Astragalus slices show structural damage such as edge chips, local corner chipping, and obvious breakage.

[0021] For each image of Astragalus membranaceus slices, defects were labeled using the least outer matrix method, and the defect categories were noted to construct the initial dataset. The labeled initial dataset was augmented using operations including rotation (−15° ~ +15°), flipping (90°), cropping (30% ~ 50%), and brightness perturbation (±20%). A random partitioning strategy was adopted, and the dataset was divided into training set, validation set and test set in a ratio of 7:2:1. The image processing module is based on an improved YOLO network structure for detecting defects in Astragalus membranaceus slices. It performs feature extraction, defect identification, and output of the results from the input Astragalus membranaceus slice image. Taking YOLOv11 as an example, the model first performs scale unification on the input Astragalus membranaceus slice image, standardizing the image size to 640×640 pixels. Next, the standardized image is input to the backbone network, which contains multiple standard convolutional layers and a C3k2 feature extraction module. Through layer-by-layer downsampling, it extracts shallow texture features and deep semantic features of the Astragalus membranaceus slices at different scales. At the end of the backbone network, a Spatial Pyramid Pooling (SPPF) module is used to fuse the feature maps at multiple scales to enhance the model's receptive field for defects of different sizes in the Astragalus membranaceus slices. Subsequently, the feature maps output from the backbone are transmitted to the neck network, which uses a Path Aggregation Network (PANet). The system employs a multi-level sampling and feature concatenation structure to fuse deep high-level semantic features with shallow high-resolution localization features, generating a fused feature map rich in contextual information. Finally, the fused feature map is input to the head module, which uses a decoupled dual-branch structure for classification and regression. The classification branch calculates the confidence level for the presence and specific defect categories of the Astragalus membranaceus slices, while the regression branch predicts the center coordinates, width, and height of the defect region's bounding box. The model output uses non-maximum suppression (NMS) to filter the predicted bounding boxes, removing redundant boxes with excessive overlap. Ultimately, the system outputs the optimal detection box location, defect category, and corresponding confidence score for the Astragalus membranaceus slice defect, completing the defect identification and localization.

[0022] The image processing module is the core of this detection method. Addressing the challenges of capturing irregular defect features in Astragalus membranaceus slices, the high false detection rate due to complex surface textures, and the stringent real-time requirements of high-throughput production lines, this module features targeted structural reconstruction and optimization of the original YOLO series network structure. The improvements to the YOLO series model are as follows, and the improved model network structure diagram is shown in the figure. Figure 4 As shown (using the YOLOv11 network structure as an example).

[0023] 1. In the main feature extraction backbone network, a linear deformable convolutional module (LDConv) is used to replace the standard convolutional layer. The structure diagram of the LDConv module is shown below. Figure 5 As shown.

[0024] To address the issue that the fixed sampling method of standard convolution is insufficient to effectively capture the features of deformed areas due to the irregular morphology of defects in Astragalus membranaceus slices, the LDConv module is introduced into the YOLO backbone feature extraction network to replace the original standard convolutional layers. LDConv adds a learnable offset to the convolutional sampling points, enabling the sampling grid to adaptively deform according to the actual feature distribution of the Astragalus membranaceus slice image content. This improvement allows the network's receptive field to dynamically fit the edge shape of mold spots or cracks instead of being limited to rectangular regions when extracting features, thus significantly enhancing the model's ability to capture features of irregular defects on the surface of Astragalus membranaceus slices.

[0025] 2. The PGLUBlock module is introduced into the feature fusion module, improving the feature fusion module structure as shown in the diagram below. Figure 6 As shown (using the feature fusion module of YOLOv11 as an example).

[0026] To significantly improve inference speed while ensuring detection accuracy, and to meet the real-time detection requirements of high-speed production lines, the feature fusion module in the YOLO series network structure has been improved in two ways: lightweighting and feature representation enhancement. Specifically, a PGLUBlock module is embedded in the feature fusion module, which consists of a partial convolution (PConv) and a gated linear unit convolution (ConvGLU).

[0027] (1). Partial convolution PConv.

[0028] PConv extracts spatial features by proportionally segmenting the input channels and performing convolution operations on only a portion of them, while preserving the original data of the unprocessed portion through identity mapping. This strategy leverages the redundancy between feature map channels, significantly reducing floating-point computation and memory access costs while retaining the main feature information of the Astragalus membranaceus slices image, thereby effectively improving the model's computational efficiency and real-time performance.

[0029] (2). Gated linear unit convolution ConvGLU.

[0030] To address the issue that the surface of Astragalus membranaceus slices contains numerous texture structures similar to defects, which can easily lead to false detections by the model, the PGLUBlock module further incorporates a ConvGLU structure after PConv. ConvGLU introduces non-linear activation capabilities through a gating mechanism, dynamically adjusting different feature channels to enhance effective features and suppress irrelevant information. Therefore, introducing ConvGLU enables the model to focus more on real defect areas and improves its ability to distinguish normal textures.

[0031] 3. Model training optimization After improving the detection model, the improved model was used to train on the surface defect dataset of Astragalus membranaceus slices. The initial parameters were optimized through multiple sets of comparative experiments. After comprehensive evaluation, the following key parameter configurations were determined: batch size of 32, training epochs of 300, initial learning rate of 0.01, final learning rate ratio of 0.01, optimization algorithm of SGD, and weight decay of 0.0005.

[0032] The results visualization module is a visualization software built on the Flask development framework, featuring four sections: a control area, a real-time detection display area, a detection result statistics area, and a defect location area. The software's visualization interface is as follows: Figure 7 As shown. The control area has buttons for model selection, save path, start detection, and end detection, which users can adjust according to the actual situation; the real-time detection display area supports real-time display of industrial camera video streams and can overlay the model output results onto the video screen in real time; the detection result statistics area will display the detection quantity information of three types of defects: cracks, mold spots, and damage in real time, which is convenient for quality monitoring and traceability analysis in the production process; the defect location area will display the category and location information of the detected unqualified astragalus slices in real time, where the location information corresponds to the coordinate system of the conveyor belt grid in the image, and the display result is the position number of the grid where each unqualified astragalus slice is located.

[0033] Figure 2 The flowchart of the Astragalus membranaceus (Huangqi) slice surface defect detection system is as follows: The system first initializes the hardware equipment, and then uses an industrial camera to capture high-definition images of the Astragalus membranaceus slices on the conveyor belt in real time. After preprocessing, the image data is input into an improved YOLO series model for inference, which quickly identifies and locates three specific defects: cracks, mold spots, and damage. The raw coordinate data output by the model is then mapped and converted into specific location numbers in the conveyor belt grid coordinate system. Finally, the system displays a video image with the model recognition results superimposed on it in the real-time detection display area of ​​the software, updates the cumulative number of the three types of defects in the statistics area in real time, and simultaneously outputs the category of unqualified Astragalus membranaceus slices and their corresponding conveyor belt grid location numbers in the defect location area.

Claims

1. A method for detecting surface defects in Astragalus membranaceus slices based on an improved YOLO series model, characterized by: The system includes a module for collecting Astragalus membranaceus slices, a module for constructing a dataset, a module for image processing, and a module for visualizing results. Its detection method includes the following steps: s1. Design and construct a hardware acquisition device for Astragalus membranaceus slices to acquire high-definition images of Astragalus membranaceus slices; s2. Based on high-resolution images of Astragalus membranaceus slices, construct an Astragalus membranaceus slice dataset for model training; s3. Improved YOLO series models trained based on Astragalus membranaceus decoction pieces dataset; s4. Develop visualization software based on the Flask framework and the trained detection model.

2. The method for detecting surface defects of Astragalus membranaceus slices according to claim 1, characterized in that: Step s1 describes designing and constructing a hardware acquisition device for Astragalus membranaceus slices to acquire high-definition images of Astragalus membranaceus slices, including: (1) Use high-resolution industrial cameras as imaging devices and configure controllable ring LED light sources to reduce ambient light interference. Adjust the brightness of the light source, the illumination angle and the camera exposure parameters to keep the acquired images high-definition and save the captured images in lossless BMP and other formats. (2) Design and use an automated conveyor belt composed of parallel square grids, each grid carrying a single piece of Astragalus membranaceus, to ensure that the Astragalus membranaceus slices are in a fixed position, with a stable spacing and no overlap during the conveying process.

3. The method for detecting surface defects of Astragalus membranaceus slices according to claim 1, characterized in that: Step s2, which involves constructing a dataset of Astragalus membranaceus slices for model training based on high-resolution images of Astragalus membranaceus slices, includes: (1) Perform quality screening on the original acquired images and remove image samples that are blurry, severely occluded, or abnormally exposed; (2) Perform image preprocessing on the screened images, including color correction and noise suppression, to ensure sample consistency; (3) Use LabelImg or other annotation tools to accurately label the Astragalus membranaceus slices, and save the labels in a TXT file format that is compatible with the YOLO series model.

4. The method for detecting surface defects of Astragalus membranaceus slices according to claim 3, characterized in that: Step (3) describes the process of accurately labeling Astragalus membranaceus slices using LabelImg or other labeling tools, which includes: 1) Defects are categorized based on surface defect characteristics of Astragalus membranaceus slices. The specific discriminant characteristics are as follows: ① Cracks: Fine, elongated linear fracture structures appear on the surface of Astragalus membranaceus slices, mostly manifested as tiny cracks distributed along the grain direction or laterally; ② Mold spots: Abnormally colored spots or patchy moldy areas that form on the surface of Astragalus membranaceus slices, usually appearing as dark, brown, or black distribution; ③ Damage: Astragalus slices show structural damage such as edge chips, local corner chipping, and obvious breakage; 2) For each image of Astragalus membranaceus slices, surface defects were annotated using the least bounding matrix method, and the defect category was noted. 3) Perform data augmentation operations on the labeled dataset using rotation, flipping, cropping, and brightness perturbation; 4) The augmented dataset is randomly partitioned into training, validation and test sets in a ratio of 7:2:

1.

5. The method for detecting surface defects of Astragalus membranaceus slices according to claim 1, characterized in that: Step s3, which involves training an improved YOLO series model based on the Astragalus membranaceus decoction pieces dataset, includes: (1) Linear deformable convolutional modules are used to replace standard convolutional layers in the backbone feature extraction network of the model, thereby enhancing the feature capture capability for irregular defects; (2) In the feature fusion module of the YOLO series models, the PGLUBlock module is introduced. The amount of computation is reduced by introducing partial convolution PConv, and gated linear unit convolution ConvGLU is used to suppress irrelevant texture features. (3) Set the training parameters, including batch size of 32, training rounds of 300, initial learning rate of 0.01, final learning rate ratio of 0.01, optimization algorithm of SGD, and weight decay of 0.0005. Use this set of parameters to train the improved model.

6. The method for detecting surface defects of Astragalus membranaceus slices according to claim 1, characterized in that: Step s4 describes building visualization software based on the Flask development framework and the trained detection model, including: (1) Initialize the Flask development framework and load the weight file of the improved model training results; (2) Construct the control area, including function buttons for model selection, save path, start detection, and end detection; (3) Construct a real-time detection display area to display the video stream captured by the industrial camera in real time and overlay the defect recognition box detected by the model; (4) Construct a defect detection quantity statistics area to count the number of defects detected in real time, including cracks, mold spots and damage; (5) Construct the defect location area, map the defect coordinates output by the model to the conveyor belt grid coordinate system, and output the grid number where the defective Astragalus membranaceus slice is located.

Images