INSA-YOLOv8: an Apple Quality Detection and Sorting Algorithm

In the processes of apple sales and apple food manufacturing, accurate identification and sorting of apple quality and damage after harvest are crucial for effectively distinguishing high-quality, normal, and damaged apples. This study proposes the INSA-YOLOv8 algorithm, which is Grounded in the optimized YOLOv8n algorithm. It has the ability to distinguish four kinds of apples: dark red high-quality, light red normal, green normal, and damaged. Substituting the C2f module in the backbone network of YOLOv8n with the iRMB module significantly enhances the model's feature extraction ability and detection accuracy. The proposed SPPF-LSKA module effectively decreases computational demands while maintaining operational efficiency, achieving concurrent improvements in detection precision and target tracking performance. The deformable selfattention mechanism (DAT) dynamically adapts the receptive field's geometry and scale, enhancing feature extraction of target regions while minimizing background distractions, thereby significantly boosting detection accuracy for distant objects. By incorporating the GSConv layer and VoV-GSCSP architecture, the network's computational framework is streamlined, alleviating redundancy in feature processing. Experimental evaluations reveal that the INSA-YOLOv8 model attains a 93.3% mAP50-95 and 99.3% mAP50, surpassing the baseline by 2.1 % and 1.8 %, respectively. Concurrently, computational load, parameter count, and GFLOPs are marginally reduced, achieving a balance between performance enhancement and resource efficiency.