Multi-Medium Image Enhancement With Attentive Deformable Transformers

Visibility challenges such as atmospheric haze, water turbidity, etc. are imposed while capturing images in various mediums like aerial, outdoor, underwater, etc. Such reduction in visibility affects the functioning of high-level computer vision applications like object detection, semantic segmentation, military surveillance, earthquake assessment, etc. The existing methods either rely on incorporating additional prior information during the training process or yield less than optimal results when analysed on images with varying levels of degradation, reason being the absence of both local and global dependencies within the extracted features. This paper presents a generalized transformer based architecture for aerial, outdoor and underwater image enhancement. We propose a novel space aware deformable convolution based multi-head self attention containing spatially attentive offset extraction. Here, the deformable multi-head attention is introduced to reconstruct fine level texture in the restored image. Additionally, we introduce a spatially attentive offset extractor within the deformable convolution to prioritize relevant contextual information. Further, we propose an edge enhancing feature fusion block for restoring the edge details in the image along with learning enriched features from multi-stream information. Finally, we propose a global context aware channel attentive feature propagator having a dual functionality of global information extraction and provision of channel attention. Comprehensive experimentation conducted on both synthetic and real-world datasets, along with thorough ablation study, showcases that the proposed approach performs optimally when compared with the existing methods on aerial, outdoor, and underwater image enhancement.