TIENet: task-oriented image enhancement network for degraded object detection

Degraded images often suffer from low contrast, color deviations, and blurring details, which significantly affect the performance of detectors. Many previous works have attempted to obtain high-quality images based on human perception using image enhancement algorithms. However, these enhancement algorithms usually suppress the performance of degraded object detection. In this paper, we propose a task-oriented image enhancement network (TIENet) to directly improve degraded object detection's performance by enhancing the degraded images. Unlike common human perception-based image-to-image methods, TIENet is a zero-reference enhancement network, which obtains a detection-favorable structure image that is added to the original degraded image. In addition, this paper presents a fast Fourier transform-based structure loss for the enhancement task. With the new loss, our TIENet enables the structure image obtained to enhance more useful detection-favorable structural information and suppress irrelevant information. Extensive experiments and comprehensive evaluations on underwater (URPC2020) and foggy (RTTS) datasets show that our proposed framework can achieve 0.5-1.6% AP absolute improvements on classic detectors, including Faster R-CNN, RetinaNet, FCOS, ATSS, PAA, and TOOD. Besides, our method also generalizes well to the PASCAL VOC dataset, which can achieve 0.2-0.7% gains. We expect this study can draw more attention to high-level task-oriented degraded image enhancement. The code and pre-trained models are available at https://github.com/ BIGWangYuDong/lqit/tree/main/configs/detection/tienet.

Automated summary

This paper proposes a task-oriented image enhancement network (TIENet) to directly improve the performance of degraded object detection by enhancing the degraded images. TIENet is a zero-reference enhancement network that adds a detection-favorable structure image to the original degraded image. The experiments and evaluations demonstrate that the proposed framework achieves significant improvements on classic detectors.