In order to achieve automated strawberry harvesting and facilitate deployment on mobile devices, this paper proposes an improved model based on YOLOv5s. The main improvements are as follows: the lightweight MobileNetV3-Small is used to replace the original CSPDarkNet53 backbone, and the network structure is further pruned to enhance the model's efficiency and adaptability for mobile devices; the CBAM attention mechanism is incorporated into different levels of the network structure to enhance the model's sensitivity to feature channels and improve the feature extraction capabilities at different levels, thereby optimizing the model's performance; to enhance the detection accuracy of small objects, a dedicated small object detection layer is added to better capture the feature information of small objects; EIOU is used instead of the traditional CIOU loss function to avoid hindering model optimization due to unclear object boundaries, further enhancing regression accuracy. The experimental results show that the mAP@0.5 and fps of the improved YOLOv5s-MNSX-C-D model in this paper are 94.3% and 48.2f/s, respectively. Compared with the YOLOv5s, Faster R-CNN, SSD and RetinaNet model, the scores of mAP@0.5 and fps have improved. The effectiveness of the improved model is verified by the ablation experiments, and a good equilibrium is obtained.
|