Pyramid Channel-based Feature Attention Network for image dehazing

Abstract

Abstract Traditional deep learning-based image dehazing methods usually use the high-level features (which contain more semantic information) to remove haze in the input image, while ignoring the low-level features (which contain more detail information). In this paper, a Pyramid Channel-based Feature Attention Network (PCFAN) is proposed for single image dehazing, which leverages complementarity among different level features in a pyramid manner with channel attention mechanism. PCFAN consists of three modules: a three-scale feature extraction module, a pyramid channel-based feature attention module (PCFA), and an image reconstruction module. The three-scale feature extraction module simultaneously captures the low-level spatial structural features and the high-level contextual features in different scales. The PCFA module utilizes the feature pyramid and the channel attention mechanism, which effectively extracts interdependent channel maps and selectively aggregates the more important features in a pyramid manner for image dehazing. The image reconstruction module is used to reconstruct features to recover a clear image. Meanwhile, a loss function that combines a mean square error loss part and an edge loss part is employed in PCFAN, which can better preserve image details. Experimental results demonstrate that the proposed PCFAN outperforms existing state-of-the-art algorithms on standard benchmark datasets in terms of accuracy, efficiency, and visual effect. The code will be made publicly available.