Distilling Object Detection Fusion of Foreground Enhancement and Spatial Frequency

Current knowledge distillation methods used in object detection often ignore the importance of global information in object detection and the quantity of foreground and background samples is unbalanced in the process of object detection training. Aiming at the above problems, a knowledge distillation method for object detection based on foreground enhancement and spatial frequency is proposed. This method uses the foreground enhancement module to introduce the attention mechanism to enhance the expression ability of the key region in the backbone feature, forcing the student model and the teacher model to focus on the same key region. The spatial frequency module guides the student model to inherit the global information contained in the teacher network by using the property of fast Fourier convolution. The losses of the neck network classifier and the regressor are calculated separately. Finally, all losses are fed into the full connection layer to obtain the final distillation loss. Experiments show that the proposed method can be applied to different object detection frameworks. Faster R-CNN, RetinaNet, and FCOS with ResNet-50 as the backbone network obtained mAP of 40.8%, 39.9%, and 42.9% on the COCO dataset, respectively, an increase of 2.4%, 2.5%, and 4.4% compared with the benchmark.