Rapid and accurate detection of multi-target walnut appearance quality based on the lightweight improved YOLOv5s_AMM...

Nut quality detection is of paramount importance in primary nut processing. When striving to maintain the imperatives of rapid, efficient, and accurate detection, the precision of identifying small-sized nuts can be substantially compromised.

We introduced an optimized iteration of the YOLOv5s model designed to swiftly and precisely identify both good and bad walnut nuts across multiple targets. The M3-Net network, which is a replacement for the original C3 network in MobileNetV3's YOLOv5s, reduces the weight of the model. We explored the impact of incorporating the attention mechanism at various positions to enhance model performance. Furthermore, we introduced an attentional convolutional adaptive fusion module (Acmix) within the spatial pyramid pooling layer to improve feature extraction. In addition, we replaced the SiLU activation function in the original Conv module with MetaAconC from the CBM module to enhance feature detection in walnut images across different scales.

In comparative trials, the YOLOv5s_AMM model surpassed the standard detection networks, exhibiting an average detection accuracy (mAP) of 80.78%, an increase of 1.81%, while reducing the model size to 20.9 MB (a compression of 22.88%) and achieving a detection speed of 40.42 frames per second. In multi-target walnut detection across various scales, the enhanced model consistently outperformed its predecessor in terms of accuracy, model size, and detection speed. It notably improves the ability to detect multi-target walnut situations, both large and small, while maintaining the accuracy and efficiency.

The results underscored the superiority of the YOLOv5s_AMM model, which achieved the highest average detection accuracy (mAP) of 80.78%, while boasting the smallest model size at 20.9 MB and the highest frame rate of 40.42 FPS. Our optimized network excels in the rapid, efficient, and accurate detection of mixed multi-target dry walnut quality, accommodating lightweight edge devices. This research provides valuable insights for the detection of multi-target good and bad walnuts during the walnut processing stage.