Hourglass Dehazing Network Based on Multi-scale Parallel Fusion
Yishu Mao, Xinman Zhang, Xingcehn Song · 2022 International Conference on Image Processing, Computer Vision and Machine Learning (ICICML) · 2022
As a common weather phenomenon, haze currently has some negative effects on advanced computer vision tasks. Image dehazing, as a significant technique, plays a crucial role in deploying downstream vision tasks. In recent years, CNN-based end-to-end models have been proved to have great advantages over traditional image enhancement-based methods in removing image haze. However, some previous works failed to effectively utilize multi-scale features and perform reasonable feature fusion, resulting in the model losing the ability to restore details and colors. In this work, we adopt hourglass network which is a mature framework to better utilize the bottom-up and top-down symmetric topology structure for feature fusion and representation, so as to better recover the details and color information of haze-free images. At the same time, for the phenomenon of relatively indirect feature fusion in hourglass network, we propose a multi-scale parallel feature fusion module to perform more direct and effective feature fusion, thereby promoting the flow of features within different scales. What’s more, for the sake of strengthening the recovery power of the network for high-frequency information, we raise a spatial-frequency domain channel attention mechanism, so as to more effectively utilize spatial and frequency information for feature representation and enhancement. Comprehensive experiments on the public RESIDE dataset confirm the effectiveness of our proposed network that can satisfactorily reconstruct more elaborate and realistic colors.