MIDNet: A Weakly Supervised Multipath Interaction Network for Image Defogging*

To overcome the overfitting problem caused by supervised defogging methods trained on synthetic datasets, some weakly supervised defogging methods trained by feeding unpaired synthetic images have been proposed. However, these methods usually express the cycle of defogging and reconstructed fog through adversarial training, without considering that this manner may be unsatisfactory for achieving the texture and detail restoration. To tackle this issue, a weakly supervised multipath interaction defogging network (MIDNet), which is trained by feeding unpaired images, is proposed. Specifically, we consider the information interaction of different generative networks. To gradually improve the quality of defogged images, MIDNet utilizes progressive paths to obtain the model parameter feature maps from different networks. The interaction of information is achieved through vertical path interaction function. Moreover, a SAMRiR module is proposed to fuse and extract features from different paths. Experiments are conducted to show that the proposed MIDNet achieves favorable performance in both qualitative and quantitative comparisons. Furthermore, we also utilize YOLOv5s to detect objects on foggy and defogged images, and experimental results demonstrate that image defogging indeed facilitates performance improvements of object detection tasks.