YOLO-GG: a slight object detection model for empty-dish recycling robot

Empty-dish recycling robot is an effective product to solve the labor shortage problem in food service. A key research point is to design a fast and accurate empty dish detection model for the robot. With the development of deep learning techniques, researchers have presented several deep learning-based models for empty-dish detection with excellent accuracy. However, to implement the empty dish detection model on an embedded device in the robot, the model should be optimized with a compact size and less calculation. To realize the optimized empty-dish detection model, this paper proposes a novel YOLO-GG model which implements the GPU-efficient Ghost module on the YOLO. In detail, the proposal adopts the G-Ghost module to replace the backbone of YOLOV4 in CSPDarknet. Meanwhile, we reduce the convolution layer in the Neck of the YOLOV4 to decrease the model parameters. The proposed model is trained on the Dish-20 dataset. Experimental results show that YOLO-GG achieves 23.4G floating point operations, 99.34% mean average precision, and 71.2 frames per second. The lower parameters and high accuracy of YOLO-GG prove the effectiveness.