Improving object detection in optical devices using a multi-hierarchical cyclable structure-aware rain removal network

Abstract

Rain streaks pose a significant challenge to optical devices, impeding their ability to accurately recognize objects in images. To enhance the recognition capabilities of these devices, it is imperative to remove rain streaks from images prior to processing. While deep learning techniques have been adept at removing rain from the high-frequency components of images, they often neglect the low-frequency components, where residual rain streaks can persist. This oversight can severely limit the effectiveness of deraining methods and consequently, the object recognition rate in optical devices such as cameras and smartphones. To address this problem, we developed a novel multi-hierarchical cyclable structure-aware rain removal network (MCS-RRN), which effectively retains the background structure while removing rain streaks, improving the object recognition rate in images. Unlike state-of-the-art approaches that incorporate wavelet transform, our network maintained the low-frequency sub-images and integrated them into a structure-aware subnetwork. We also transferred low-frequency structural information to detail enhancement sub-networks to enhance detailed information and facilitate convergence; this enhanced the capability of our network to eliminate rain streaks in high frequency. In summary, we used a structure information blending module and inverse wavelet transform to fuse derained low-frequency sub-images and achieve rain removal while improving the object recognition rate with the combination of YOLO. Experimental results demonstrated that our method significantly enhances the object recognition rate in images.