Air infrared small target local dehazing based on multiple-factor fusion cascade network

Infrared (IR) imaging, an import method of target monitoring, suffers great imaging quality degradation in poor weather conditions such as haze, fog, and smog. This will greatly affect the ability of detecting and identifying for targets with far distance. For visible-light imaging, an image processing technique named dehazing has been developed in the past several years. However, these dehazing methods for visible light can hardly be used to IR picture dehazing directly, due to the natural difference between IR and visible-light images. In this paper, an IR image dehazing algorithm based on multiple-factor fusion cascade network (MFFCN) is proposed, which includes multi-patch image encoder, multi-channel feature enhancement module and multi-level feature fusion module to directly remove the haze. Specifically, a multi-patch image encoder aggregating features from multiple patches of image to improve the response for different levels of haze in different regions and a multi-channel feature enhancement module can provide interactions of cross-feature and enrich the diversity of learned representations. A multi-level feature fusion module is integrated to calculate the importance weights of different network levels, and then, the calculated weights are utilized to fuse the corresponding feature. The experimental results showed that our MFFCN achieves the highest performance with both visual perception and evaluation metric, when compared with previous state-of-the-art methods. Particularly, for thick hazing scene, a greatly contrast enhancement for the local area around target has been achieved. In addition, the effectiveness for ground background image hazing and halation removing is demonstrated, which can also obtain superior performance.