YOLO-OR: a lightweight cross-stage object detection model for dish recycling robot

With the development of deep learning, object detection technology has been widely used in various areas. However, the large size of the model and the high computation cost limit the deployment on the devices of the robot. To address these problems, this study proposes a slight cross-stage object detection model YOLO-OR for the dish recycling robot. In detail, the cross-stage octave convolution combines with the residual structure to build a novel convolution block ORBlock, improving feature extraction capability through cross-stage convolution operations. Moreover, ORBlock significantly reduces the parameters via splitting the traditional convolution into two-dimension. In particular, the four outputs of the backbone’s last two ORBlocks are merged and transmitted to the YOLO neck for feature fusion. Based on the Dish-20 dataset, YOLO-OR achieves excellent performance with 99.56% mAP, 189.42 FPS, 4.64 G FLOPs and 3.29 MB parameters. Meanwhile, the performance of YOLO-OR outperforms other state-of-the-art YOLO models. These demonstrate the effectiveness of the proposal.