Haze is the result of the interaction between specific climate and human activities. When observing objects in hazyconditions, optical system will produce degradation problems such as color attenuation, image detail loss and contrastreduction. Image haze removal is a challenging and ill-conditioned problem because of the ambiguities of unknownradiance and medium transmission. In order to get clean images, traditional machine vision methods usually use variousconstraints/prior conditions to obtain a reasonable haze removal solutions, the key to achieve haze removal is to estimatethe medium transmission of the input hazy image in earlier studies. In this paper, however, we concentrated on recoveringa clear image from a hazy input directly by using Generative Adversarial Network (GAN) without estimating thetransmission matrix and atmospheric scattering model parameters, we present an end-to-end model that consists of anencoder and a decoder, the encoder is extracting the features of the hazy images, and represents these features in highdimensional space, while the decoder is employed to recover the corresponding images from high-level coding features.And based perceptual losses optimization could get high quality of textural information of haze recovery and reproducemore natural haze-removal images. Experimental results on hazy image datasets input shows better subjective visualquality than traditional methods. Furthermore, we test the haze removal images on a specialized object detection network-YOLO, the detection result shows that our method can improve the object detection performance on haze removal images,indicated that we can get clean haze-free images from hazy input through our GAN model.
展开▼
