Multi-Feature Fusion ECANet-Based Adaptive Dehazing Algorithm for Aerial Images

Abstract

Aiming at the problems of blurred details, insufficient utilization of multi-scale features, brightness shift and color distortion in aerial hazy images, this paper proposes a multi-feature fusion attention-based adaptive dehazing network (MFFAMA-Net), which is composed of three cascaded modules, feature extraction, multi-feature fusion attention and adaptive brightness correction. The feature extraction module preserves shallow edge and texture details, and adopts three-branch parallel multi-scale convolutions to enlarge the receptive field level by level while keeping the number of channels constant, with skip connections linking the shallow and deep information. In the multi-feature fusion attention module, global max pooling (GMP) is introduced in parallel into the channel attention of ECA-Net without dimensionality reduction, and is combined with global average pooling (GAP) to form a dual-path encoding, so as to realize channel recalibration and adaptive fusion of shallow and deep features. The adaptive brightness correction module predicts pixel-wise and channel-wise Gamma parameters through the network to adaptively correct the brightness and color of the restored result. The network is optimized in an end-to-end manner with a multi-constraint joint loss function composed of a pixel-level reconstruction loss, a perceptual loss and a color consistency loss. Experimental results show that on the RESIDE dataset the proposed method outperforms the mainstream dehazing methods in objective metrics such as PSNR and SSIM, and the information entropy also indicates that the restored images contain richer details. Validation on the real aerial hazy images selected from HazyDet shows that the proposed method still maintains a good visual restoration effect in real scenes and has favorable practical applicability. In addition, the inference time of the proposed method for a single image is 0.216 s, which can meet the requirement of fast image dehazing.