Single Image Dehazing based on Multi-Scale Feature Fusion Under Adam-Optimization

To address the issues of insufficient highlight extraction in existing dehazing algorithms and an abundance of image data, a novel approach based on multi-scale feature fusion is put forth. First, U-Net’s fundamental convolutional layer is enhanced with updated fully connected residual blocks to reduce computational complexity. To extract minute detail from the images, a self-convolution module that makes use of self-attention mechanisms is also presented. Moreover, scale-gated units are used to merge feature maps from various levels in order to maximize feature reuse and reduce information loss. A unique loss function is presented that combines minimum absolute error and multi-scale structural similarity to enhance the recovered images’ subjective impression. Synthetic haze datasets are used for experimental assessments, which show notable gains over alternative neural network architectures. In particular, the suggested technique improves the dehazed images’ signal-to-noise ratio by 18.33% and multi-scale structural similarity by an average of 4.31%. These findings show that the suggested method successfully reduces problems such strong edge artifacts, halo effects, and color distortion that are frequently seen in fuzzy photos. Moreover, following haze reduction, the recovered photos show good subjective recognition, underscoring the usefulness of the suggested technique.