Using filter pruning-based deep learning algorithm for the real-time fruit freshness detection with edge processors

In the context of edge devices used in horticulture and the fruit industry, traditional high-precision models prove to be computationally intensive and challenging to deploy on terminals with limited computing resources. Therefore, in this work, we propose a lightweight target detection algorithm, called Fruit-Yolo, applied to edge farming devices where there are insufficient hardware resources and weak computational power to achieve classification and localization of fruit. More specifically, we employ the PPYOLO-Tiny model as our benchmark, where we undertake the redesign of both the backbone and FPN network components of the model. Additionally, we adapt the anchor framework to better suit the requirements of fruit detection, ultimately resulting in the creation of our streamlined network (Fruit-Yolo). Then, the filtered pruning based on geometric median algorithm is used to prune the model. The test results showed that the number of parameters of model after pruning was reduced by 20%, and the mAP decreased by only 0.2%. To verify the effectiveness of the proposed method, five algorithms including theYOLOv5s, SSD 300, MobileNetv3 and Faster R-CNN were compared. The comparison results show that the Fruit-Yolo model has the smallest size (3.85 MB), high accuracy (98.6%) and fast detection speed (32.77FPS). The results indicated that the proposed method can provide a technical reference for the deployment of deep learning models in intelligent agricultural machines at the edge.