DG-YOLOX: A Novel Single-Stage CNN-based Method for Workshop Object Detection

To realize workshop object detection, this paper firstly builds a workshop object detection dataset composed of computer numerical control machine tools and robot arms and proposes DG-YOLOX based on YOLOX-Tiny. Then, to improve the quality of upsampling features of the model, an upsampling module based on the Warp-Up operation is proposed. Next, the deformable grid decoupled head is proposed to increase the flexibility of sample selection during training. It uses grid deformation and deformable convolution to help the model learn multi-scale information. To avoid excessive grid offset, the grid offset mechanism is only applied in the last two detection layers of the model. Additionally, to make the grid offset mechanism bring more gains to the model, this paper proposes a feature fusion module across tasks and layers to enhance the classification feature information and the regression feature information in different feature layers. Finally, the validity of the model is additionally validated using the VOC2012 dataset. In the VOC2012 dataset, compared to baseline, the performance is improved by 11.5%, 10.9% and 7.9% respectively for AP@50, AP@75 and AP@50:95. In the workshop object detection dataset, AP@50 is improved by 15.6% compared to the baseline. The comparison with other mainstream detectors also demonstrates the validity of the proposed model.