SIP-Net: A Shift Interactive Perception Network for Single Image Dehazing

Learning-based methods have demonstrated their effectiveness in image dehazing tasks. Recent variants of Swin Transformer have made significant progress in image dehazing. However, they limit the self-attention (SA) calculation within each window, which reduces the image dehazing performance. To overcome this question, we propose a Shift Interactive Perception Network (SIP-Net) in this paper for image dehazing. Specifically, we first use a Shift Convolution Block (SCB) to effectively extract local information of image features. Next, to the interdependence of channel features for more powerful feature relevance learning, we embed a Second-order Attention Block (SAB) into SIP-Net. Following the SAB, we propose an Interactive Perceptual Block (IPB) involving a three-branch architecture. To capture the complex dependencies of the whole image, the upper branch adopts a recurrent cross-attention block, which can effectively model the contextual information of all pixels. The bottom branch extracts feature maps while increasing their edge details using a simple and efficient encoding-decoding architecture. To improve the representation, the middle branch adaptively updates the characteristics based on the outputs of the top and lower branches. Extensive evaluations on synthetic and real-world datasets demonstrate that our SIP-Net outperforms other state-of-the-art dehazing methods.