Robust detection of dehazed images via dual-stream CNNs with adaptive feature fusion

摘要

Image dehazing is common post-processing in automatic driving and video surveillance, which can improve image visual quality. However, it might also be used as an image forgery that is difficult to be perceived by naked eyes. Though image dehazing has attracted wide attention, there are still no works specially designed for this kind of forgery detection. By making extensive experiments and preliminary analysis, we observe that a dehazed image easily loses its illumination consistency that can be captured by inverse-intensity chromaticity (IIC) transformation. IIC is a transformed color space that well represents image illuminance map. In this work, a dehazing detection network (DDNet) is proposed to distinguish dehazed images from natural haze-free images. The proposed DDNet accepts RGB images and IIC images as inputs, which are fed into the backbone network, namely EfficientNet-B0, to learn features, respectively. To effectively fuse the learned RGB and IIC features, we also propose an adaptive feature fusion method, thereby improving detection capability. In addition, we build a dehazed image dataset, which includes 11432 pairs artificial, 367 pairs natural hazy images and their corresponding dehazed images, as the benchmark. Experimental results prove that the proposed DDNet achieves desirable detection accuracies and satisfactory robustness.