DFL-Net: Effective Object Detection via Distinguishable Feature Learning

The one-stage anchor-based approach has been an efficient and effective approach for detecting objects from massive image data. However, it neglects many distinguishable features of objects, which will lower the accuracy of object detection. In this paper, we propose a new object detection approach that improves existing one-stage anchor-based methods via a Distinguishable Feature Learning Network (DFL-Net). DFL-Net integrates distinguishable features into the learning process to improve the accuracy of object detection. Notably, we implement DFL-Net by a full-scale fusion module and an attention-guided module. In the full-scale fusion module, we first learn the distinguishable features at each scale (layer) and then fuse them in all layers to generate full-scale features. This differs from prior work that only considered one or limited scales and limited features. In the attention-guided module, we extract more distinguishable features based on some positive or negative samples. We conduct extensive experiments on two public datasets, including PASCAL VOC and COCO, to compare the proposed DFL-Net with several one-stage approaches. The results show that DFL-Net achieves a high mAP of 83.1% and outperforms all its competitors. We also compare DFL-Net with three two-stage algorithms, and the results also suggest the superiority of DFL-Net.