Research on YOLOv5 Lightweight Algorithm for Camellia Oleifera Fruit Recognition

Due to the large number of parameters and high computational complexity, it is difficult to deploy a generalized detection and recognition model directly on mobile. In order to solve the difficulty of lightweight mobile deployment and enhance the recognition ability of Camellia oleifera fruit on mobile devices, a lightweight and real-time detection model is proposed. First, a depth-separable convolution is introduced to scale down the traditional YOLO v5s network, and the unimportant channels are pruned using model sparsity training and batch normalized scaling factor distribution condition. Finally, the bounding box regression loss function is improved to EIoU loss function to obtain more accurate box localization to enhance the model recognition ability. The experimental results show that the improved model can be effectively lightweight while maintaining the model performance. Compared with the traditional YOLO v5s, the mAP of the improved algorithm is increased by 0.2 percentage points, while the weight of the improved model is 35.68, which is 8.35MB less than the original model, and the FPS is 35.68, which is able to satisfy the requirements of the deployment of mobile and embedded devices, respectively.