Task-driven Deep-Learning-based Image Enhancement

With the popularization of digital image applications,the technique of digital im-age enhancement is becoming more and more important.In recent years,the continuous improvement of deep learning technique also promotes the image enhancement algo-rithms.Nevertheless,it still faces some challenges now,such as improving efficiency of deep networks,generalizability in real-world data,task-oriented optimization,etc.In this thesis,we focus on exploring and discussing the solutions to these challenges in practical image enhancement problems.First,the quality of image color influences viewing experience a lot.Due to that global harmonious perception and local details should be well-considered in a sin-gle model simultaneously,we propose a coarse-to-fine color enhancement algorithm.Within our algorithm,the color enhancement process is divided into channel-wise global enhancement and pixel-wise local refinement performed by two cascaded Convolu-tional Neural Networks(CNNs).In channel-wise enhancement,our model predicts a global linear mapping for RGB channels of input images to adjust global style.In pixel-wise refinement,we learn a refining mapping using residual learning for local adjust-ment.Further,we adopt a non-local attention block to capture the long-range dependen-cies from local information for subsequent fine-grained local refinement.Channel-wise enhancement generates coarse enhanced results with global harmonious perception and pixel-wise refinement adjusts local details to inhibit local artifact.The experiments in-dicate the effectiveness of two proposed enhancers,and the performances of proposed method outperform state-of-the-art works in quantitative or qualitative comparisons.Second,the distortion in real-world scene is sophisticated,it easily occurs condi-tions that several different distortions are overlaid together,which is called combined distortion.We propose a Dual Prior Learning(DPL)algorithm for combined-distortion image enhancement,inspired by Deep Image Prior(DIP).The proposed algorithm con-tains image prior network and distortion prior network.The image prior network can generate a clean image from random noise through inherent prior in given distorted im-age.The distortion prior network can predict distortion type and distortion level from distorted image,which are used for converting the generated clean image into corre-sponding distorted image.In this way,the constraints from generated image to the dis-torted one can be established.Furthermore,adversarial learning is applied to keeping the distribution of generated images close to the distribution of clean images,which is aimed to improve subjective visual quality.The experiments indicate the effectiveness of proposed algorithm,and the performances of proposed model outperform DIP on unknown combined distortion enhancement in quantitative or qualitative comparisons.In addition,distortion-perception-semantic quality tradeoff generally exists in deep-learning-based enhancement algorithms.In view of this observation,we further discuss task-oriented optimization for deep-learnng-based color enhancement and com-bined distortion enhancement.Concretely,we discuss the influence of perceptual loss and adversarial learning for color enhancement and the influence of classification loss from recognition model for combined distortion enhancement.