Research Of Lightweight Algorithm And System For Low-light Image Enhancement In Semi-supervised Learning

Digital image is an important carrier of information transmission,which is widely used in entertainment,security,computer vision and other fields.Restricted by insufficient illumination at night,images and videos often suffer from a series of problems.In addition,existing high-level tasks in computer vision rely heavily on highquality image inputs.Therefore,an effective way to enhance low-light images is urgently needed.With good generalization and nonlinear expression ability,the current low-light image enhancement method is based on the deep learning.Those supervised methods suffered from the problem of data collection and domain adaptation.Therefore,the current work has turned to unsupervised methods.Lacking strong constraints in model optimization,the results of unsupervised method also seem unsatisfactory.Based on the above problems,semi-supervised method combined with small supervised dataset is an alternative way.In addition,the low-light image enhancement task is sophisticated.Self-paced learning can learn samples from easy to difficult,and help model to achieve a better local optimum which is also an alternative way.For these reasons,this paper conducts research on low-light image enhancement algorithms under the framework of self-paced learning and semi-supervised learning.From the perspective of engineering practice,a lightweight real-time enhancement system is further designed for those Android devices by using model compression techniques.The main contribution of this paper is as follows:(1)With the combination of supervised and unsupervised samples,this paper proposes the first jointly training semi-supervised low-light image enhancement algorithm.On the basis of DRBN,an unsupervised image selection part and a semisupervised low-light image enhancement part are applied to overcome the problem of lacking extra same domain unsupervised images and the separated supervised and unsupervised modules.In the first part,a scoring mechanism based on the QTP theory is used to score unsupervised low-light images,images with lower score are selected as the extra same domain unsupervised images for low-light image enhancement task.In the second part,we extend the Mix Match based semi-supervised classification algorithm into its semi-supervised regression version,and utilize recursive band learning(RBL)which is the first stage of DRBN as the supervised learning part of our model to complete the task.Comprehensive experiments demonstrate the effectiveness of the proposed method.(2)From the perspective of sample learning order,this paper further proposes an unsupervised deep self-paced regression framework for low-light image enhancement task.The self-paced regression framework consists of two parts: the regression learning process with curriculum and the curriculum generation process based on the learned regression model.Through the repeated action of the regression learning process and the curriculum generation process,the pseudo-regression dataset with high confidence in the same domain is selected from easy to difficult as a new training dataset.By using the pseudo-regression dataset with high confidence,a more accurate regression(such as low-light image enhancement)models can be learned.Comprehensive experiments demonstrate the effectiveness of the proposed method.(3)From the perspective of engineering practice,the existing deep learning algorithms rely heavily on the hardware and software of the PC.In addition,there is no relevant research on placing low-light image enhancement algorithms on devices other than PC.Considering the popularity of mobile devices,this paper designed a low-light image enhancement system based on Android devices to achieve real-time low-light image enhancement by using the compressed lightweight enhancement model.