A Method for Quality Grading and Endogenous Impurity Detection of Walnut Kernels Based on Lightweight Deep Learning

This paper proposes a lightweight detection model YOLO-WQ based on improved YOLOv11n to meet the multiple requirements of high precision, lightweight, and real-time for quality grading and endogenous impurity detection in walnut kernel processing. Traditional methods are difficult to effectively identify small targets such as walnut kernels and their impurities with complex surface textures and similar shapes, while existing deep learning models suffer from high computational resource consumption and deployment difficulties. This study conducts triple optimization starting from the model structure: ShuffleNetV2 is used as the backbone network, significantly reducing the number of parameters and computational burden; Design a Slim neck neck structure that integrates GSConv and VoV-GSCSP to enhance multi-scale feature fusion capability; Introducing SEAM attention mechanism to improve the perception accuracy of walnut groove details and occluded areas. Experiments have shown that YOLO-WQ achieves success on self built datasets [email protected] At 90.6%, the accuracy and recall rates reached 84.8% and 87.4%, respectively, which were 2.5%, 4.2%, and 3.7% higher than the original model. The computational load (FLOPs) decreased by 44.4%, the model volume was reduced by 20%, and the inference speed reached 157.09 fps. Further testing on multiple publicly available agricultural product datasets has confirmed its excellent generalization performance. This model combines high precision and efficiency, and is suitable for real-time quality monitoring and impurity detection tasks in actual walnut processing production lines.