Deep Learning Methods For Shadow Removal And Haze Removal

Shadow and haze can affect the quality of digital images and the performance of computer vision algorithms.Therefore,single image shadow removal and haze removal plays a vital role in computer vision.Existing research on image shadow removal and haze removal failed to explore the connection between shadow and haze.Deep learning methods for image shadow understanding only focus on shadow detection or shadow removal,but never utilize the correlation between the two tasks.For image haze removal,the topology of existing dehazing backbone networks is not good enough.In response to above problems,this paper carried out the following research:(1)This paper introduces the relationship and difference between shadow and haze from the perspective of their formation theory.At the same time,the theoretical analysis of color-lines is given,and the relationship between the shadow image and the haze image can be deeply understood through the change of color-lines.Shadow removal and haze removal are further explored by studying the similar methods based on color-lines.These explorations have discovered potential connections between the two unrelated tasks from theoretical and methodological levels,allowing us to understand shadow and haze in the image deeply.(2)A novel STacked Conditional Generative Adversarial Network(ST-CGAN)for jointly learning shadow detection and shadow removal is proposed,by making full use of the correlation between the two tasks.Extensive experimental results consistently show the advantages of ST-CGAN over several representative state-of-the-art methods on some publicly available datasets.At the same time,we construct the first large-scale image shadow triplets dataset(ISTD),which contains 1870 triplets(shadow image,shadow mask image,and shadow-free image)under 135 scenes.(3)A new dehazing backbone network,All-in-one Mixed Link Network for Single Image Dehazing(AMN),is presented based on the "dense topology" of mixed link net-works(MixNet).At the same time,we also introduce a discriminator for the dehazing backbone network to form an adversarial model.Extensive experimental results show the advantages of AMN over many representative dehazing methods on several public datasets and our synthesized one.