Context Convolution Dehazing Network With Channel Attention

Abstract

Fog and haze weather conditions lead to deterioration of image visual quality. Under these conditions, advanced image processing tasks such as object detection and image segmentation are difficult to perform. To solve related problems, in this paper we propose an end-to-end context dilated convolution dehazing network with channel attention to return to a clear image from a haze image. This model uses context dilated module extracts the multi-scale scene information in the haze image, which can better maintain the original color of the image while removing the haze. The channel attention enables the model to separate the importance of features and boosts the model’s power to adapt to different input scenarios. In the training phase, the model uses contrastive learning to distinguish the potential difference between the haze picture and the clear picture, helping the model to better renewal the clear image. The results of contrastive tests with existing methods indicate the proposed method has excellent dehazing performance.