ASD-YOLO: a lightweight network for coffee fruit ripening detection in complex scenarios

Coffee is one of the most popular and widely used drinks worldwide. At present, how to judge the maturity of coffee fruit mainly depends on the visual inspection of human eyes, which is both time-consuming and labor-intensive. Moreover, the occlusion between leaves and fruits is also one of the challenges. In order to improve the detection efficiency of coffee fruit maturity, this paper proposes an improved detection method based on YOLOV7 to efficiently identify the maturity of coffee fruits, called ASD-YOLO. Firstly, a new dot product attention mechanism (L-Norm Attention) is designed to embed attention into the head structure, which enhances the ability of the model to extract coffee fruit features. In addition, we introduce SPD-Conv into backbone and head to enhance the detection of occluded small objects and low-resolution images. Finally, we replaced upsampling in our model with DySample, which requires less computational resources and is able to achieve image resolution improvements without additional burden. We tested our approach on the coffee dataset provided by Roboflow. The results show that ASD-YOLO has a good detection ability for coffee fruits with dense distribution and mutual occlusion under complex background, with a recall rate of 78.4%, a precision rate of 69.8%, and a mAP rate of 80.1%. Compared with the recall rate, accuracy rate and mAP of YOLOv7 model, these results are increased by 2.0%, 1.1% and 2.1%, respectively. The enhanced model can identify coffee fruits at all stages more efficiently and accurately, and provide technical reference for intelligent coffee fruit harvesting.