AGN: Adversarial Grafting Network for Few-Shot Visual Haze Classification

Haze affects the accuracy of computer vision algorithms, so identifying the haze density is particularly important. Existing works are based on extracting practical haze features from abundant real-world images, and obtaining these labeled images requires enormous manual costs. In addition, the accuracy of haze classification still needs to be improved to be better applied in image processing tasks. To address this issue, this paper develops a dual-channel generative framework for model enhancement, termed adversarial grafting network (AGN), which aims to graft realistic haze features onto haze-free images for visual few-shot haze classification. The critical procedure in grafting is to separate haze-style related features from content features representing invariant semantic content. The performance of classifiers trained on synthetic images is effectively enhanced. To better study the field of hazy vision, we build the multiple source foggy image dataset (MSFID) with detailed density discrimination. Experiments were conducted on MSFID regarding performance comparison, few-shot learning, ablation studies, and parameter sensitivity. The results demonstrated that images appropriately grafted with haze-style features are adequate for visual few-shot haze classification, and the proposed AGN consistently improves the performance of the state-of-the-art deep learning backbones.