Abstract: To achieve real-time detection of strawberry ripeness in complex environments, a lightweight object detection model YOLOv5s-DCC based on an enhanced YOLOv5s architecture was proposed, aiming to enhance both accuracy and computational efficiency. Depthwise separable convolutions were embedded within the backbone structure to reduce computational complexity. The neck section incorporated CARAFE upsampling to enhance the capture of minute features, while the detection head integrated CBAM attention mechanisms. Dynamic weighting improved feature selection performance, resulting in an optimized model balancing detection accuracy and lightweight efficiency. The refined model size was only 13.9 MB, achieving precision, recall, and mean average precision (mAP) of 93.4%, 92.7%, and 95.3%, respectively. Compared to the original YOLOv5s model, precision, recall, and mAP improved by 1.3, 0.9, and 1.6 percentage points, respectively, while reducing parameters by 0.5×10⁶. Compared to mainstream lightweight models such as YOLOX-s, YOLOv7-tiny, and YOLOv8s, YOLOv5s-DCC deliverd the best overall performance. It could meet the real-time strawberry ripeness detection requirements of agricultural harvesting robots in complex environments.

Key words: strawberry ripeness detection, lightweight model, depth separable convolution, CBAM attention mechanism, CARAFE upsampling