Dense spatially-weighted attentive residual-haze network for image dehazing

Haze severely affects computer vision algorithms by degrading the quality of captured images and results in image data loss. With several available approaches for dehazing, single image dehazing is most preferred and challenging. We proposed a Dense Spatially-weighted Attentive Residual-haze Network (DSA Net), a novel end-to-end Encoder-decoder architecture to learn the residual haze layer between the hazy and haze-free image. We use encoder-decoder blocks with multiple skip connections to improve feature propagation. Feature Learning block uses a novel Residual Inception fused with Attention (RIA) block to learn the complex non-linearity from features extracted from the encoder part. Learning residual image is more straightforward than the whole haze-free image, and it improves the ability of the network to estimate the haze thickness accurately. DSA Net learns this less complex residual-map from the hazy input image and subtracts it from the input to obtain the dehazed images. Detail ablation study shows the effectiveness of different modules used in our architecture. Experiment results on different haze conditions demonstrate that our method shows significant improvement over other state-of-the-art methods.