Research On Key Technologies Of Image Enhancement And Recognition In Low-light Conditions

Due to limited photons and inevitable noise,images captured in low-light conditions suffer from very poor visibility caused by low contrast,color distortion and significant measurement noise.It is hard to extract information from such images effectively,while the spatially unevenly distributed illumination and noise make things even harder.It is of great challenge for real-world applications of image recognition and enhancement algorithms in computer vision systems.The problem of low-light image enhancement can be recast as the Retinex decomposition problem.It is an ill-posed inverse problem,and the low SNR of the input low-light image further aggravates the ill-posedness.Based on the inherent connections between noise and illumination,this dissertation firstly studies under the settings of supervised learning with training data and unsupervised learning without training data,how to perform robust Retinex decomposition in the presence of significant noise.Then,inspired by the inconsistency of the purpose of enhancement and recognition,as well as the low efficiency of the existing unsupervised learning model training,this dissertation studies how to obtain a low-light recognition model with high training efficiency and strong generalization ability.This dissertation aims to improve the visibility and quality of low-light images so as to improve the robustness of subsequent visual recognition algorithms to lighting variation,such that the computer can “see”clearly and accurately even in low-light conditions.The research results of the dissertation are summarized as follows.1.Based on the prediction of pixel-wise operators defined in the edge-aware bilateral space,this dissertation proposes a deep-learning-based approach for low-light image enhancement by estimating the illumination and the noise simultaneously.It can accurately predict the desired normal-light images in the presence of significant spatially-varying measurement noise.The proposed method showed its advantage over other methods when processing images captured in extremely low-light conditions.2.Based on the deep image priors captured by the architecture of convolutional neural networks and an illumination-adaptive self-supervised denoising module,this dissertation proposes an unsupervised method for low-light image enhancement.It exploits the powerful modeling capability of deep learning while requiring no training samples except the input image itself.The proposed method competes favorably with supervised-learning-based methods in certain scenarios,with strong robustness to complex noise in low-light images.3.Inspired by the idea of multi-exposure,this dissertation proposes an end-to-end learning framework that jointly perform image enhancement and recognition for the task of low-light face detection.It can not only effectively inhibit non-uniform illumination and noise issues so as to significantly improve face detection performance,but also be flexibly coupled with different off-the-shelf face detectors.The proposed method outperforms previous state-of-the-arts by a significant margin on the low-light face detection benchmark.4.This dissertation makes an attempt to solve the problem of low model training efficiency for low-light semantic segmentation with adaptive learning rate as the starting point.This dissertation investigates the potential of a convex optimization method with low computational complexity and fast convergence for training neural networks.The method is generalized to adaptive learning rate of mini-batch gradient descent,with strong motivation from related convergence analysis.The proposed method has its advantages on both learning speed and generalization performance over other available methods.