Visibility Enhancement in Surveillance images using Deep Multi-scale Feature Fusion

Visibility degradation is expected in adverse weather phenomena such as fog, mist, and haze. Object detection and identification in the surveillance feed are challenging in these weather conditions. Outdoor images can naturally be degraded, or one can intentionally degrade using smoke or bright light to hide identification. Haze, fog, and smoke are pixel-based degradation and can vary based on their thickness and distribution property in an image. In our proposed work, we explore the possibility of identifying and removal hazy pixels using the benefits of multi-scale feature-fusion and In-scale feature progression. We proposed a learning-based end-to-end network for single image dehazing. Our proposed architecture consists of three different modules: (1) Coarse Feature-fusion, (2) Fine Feature-fusion, and (3) Reconstruction module. The Coarse feature-fusion module learns broad contextual information, and the Fine feature-fusion module refines the coarse features by focusing on the channel and pixel-based information. Multi-scale feature fusion is used both in the coarse and fine module to benefit the network stage from the previous stage’s output. Extensive experimental results suggest that the proposed approach outperforms other state-of-the-art methods on synthetic homogeneous and non-homogeneous haze data and improves object detection and identification accuracy.