Small Object Detection Based on Super-Resolution Enhanced Detection Network

Abstract

Object detection is one of the most important fields in computer vision and plays a key role in various practical application scenarios. Although most of the current mainstream object detectors have achieved outstanding results in the detection of large and medium-sized objects, the detection accuracy of small objects lacking sufficient detailed information is still unsatisfactory. To solve this problem, we propose a super-resolution enhanced detection network (SREDet). In the SREDet, we improve and integrate ESRGAN and YOLOv3 to make the joint loss of the entire detection network can assist in optimizing the generator, so that the generated super-resolution images are more conducive to the task of small object detection. At the same time, the method of edge extraction and enhancement is adopted in our network to reduce the impact of background noise on small object detection, while its loss function is also improved by us. In addition, we propose an image segmentation and mapping module to adapt the input size of super-resolution images and detectors. Finally, we completed the whole object detection network with an improvement of NMS algorithm. Our experiments are based on the COCO dataset, and the results demonstrate the effectiveness of each module. Compared with some classic object detectors, our network has a significant improvement in the detection performance of small objects.