MLAFF-Net: Multi-level Attention-based Feature Fusion Network for Single Image Dehazing

Abstract

We propose a convolution neural network to recover a dehazed image titled Multi-level Attention-based Feature Fusion Network (MLAFF-Net). MLAFF-Net consists of feature extraction blocks, pixel attention, feature fusion blocks, and mixed convolution attention mechanism. Feature extraction block is employed to extract the features. MLAFF-Net has ability to focus on significant features employing pixel attention mechanism. Feature fusion block fuses and refines the significant features of various levels for next fusion block. The accurate estimation of the kernel may recover a sharp image. Moreover, MLAFF-Net has ability to acquire both the high-level and low-level significant features, reduce the feature redundancy and boost the further internal feature representations employing the mixed convolution attention module. Further, multi-level supervision learning method is employed to compute the loss at various resolution levels. The experimental findings show that MLAFF-Net exhibits outstanding performance when compared to existing single image dehazing methods for both synthetic and real-world images.