DENet: Detection-driven Enhancement Network for Object Detection under Adverse Weather Conditions

Qingpao Qin, Kan Chang, Mengyuan Huang, Guiqing Li; Proceedings of the Asian Conference on Computer Vision (ACCV), 2022, pp. 2813-2829

Abstract


Recently, the deep learning-based object detection methods have achieved a great success. However, the performance of such techniques deteriorates on the images captured under adverse weather conditions. To tackle this problem, a detection-driven enhancement network (DENet) which consists of three key modules for object detection is proposed. By using Laplacian pyramid, each input image is decomposed to a low-frequency (LF) component and several high-frequency (HF) components. For the LF component, a global enhancement module which consists of four parallel paths with different convolution kernel sizes is presented to well capture multi-scale features. For HF components, a cross-level guidance module is used to extract cross-level guidance information from the LF component, and affine transformation is applied in a detail enhancement module to incorporate the guidance information into the HF features. By cascading the proposed DENet and a common YOLO detector, we establish an elegant detection framework called DE-YOLO. Through experiments, we find that DENet avoids heavy computation and faithfully preserves the latent features which are beneficial to detection, and DE-YOLO is effective for images captured under both the normal condition and adverse weather conditions. The codes and pre-trained models are available at: https://github.com/NIvykk/DENet.

Related Material


[pdf] [code]
[bibtex]
@InProceedings{Qin_2022_ACCV, author = {Qin, Qingpao and Chang, Kan and Huang, Mengyuan and Li, Guiqing}, title = {DENet: Detection-driven Enhancement Network for Object Detection under Adverse Weather Conditions}, booktitle = {Proceedings of the Asian Conference on Computer Vision (ACCV)}, month = {December}, year = {2022}, pages = {2813-2829} }