Research On Low Light Natural Image Enhancement Algorithm

As one of the crucial carriers of information,images act an indispensable role in national key needs and human daily life.Yet,due to factors such as lighting conditions and shooting equipment,the collected low-light images cannot efficiently transmit information.These images are often accompanied by a series of problems such as low contrast,color distortion,noise interference and lack of details,which not only affect human visual perception,but also limit subsequent processing and applications based on computer vision.Therefore,it is of great theoretical significance and application value to carry out research on low-light enhancement algorithms.In recent years,deep learning-based low-light image enhancement algorithms have made massive progress,but there are still problems and challenges.On the one hand,most algorithms do not consider or effectively deal with the noise interference in low-light images.On the other hand,due to the difference in the distribution of synthetic data and real data,the generalization performance for real images needs to be improved.In response to the above problems,the thesis takes low-light images as the research object and uses deep learning to study the problems in low-light image enhancement.The main contents and contributions are as follows:1)An attention-based multi-branch network for low-light image enhancement algorithm is designed.Images acquired in low-light conditions have low visibility,and due to the influence of many factors,the illuminance intensity in the scene will vary with the spatial region.At the same time,there are also varying degrees of noise interference in low-light regions.Based on this,an algorithm of attention-based multibranch network is designed for low-light image enhancement.The algorithm is based on Retinex theory.First,the input low-light image is decomposed into reflectance component and illumination component,and then the reflectance and illumination are adjusted respectively in the two sub-networks named reflectance restoration module and lighting adjustment module,and the adjusted components are corresponding multiply to obtain the final enhanced image.A lightweight and effective attention block is added to the reflectance restoration module,and different regions of the reflecance component are given different attention by introducing the illumination information of the original image,so that the backbone network of the module can more effectively restore the reflectance component and suppress noise.In addition,a lightweight convolutional neural network is used to adjust the illumination component.Finally,experiments were carried out on different low-light image datasets,and compared with other algorithms,which verified the effectiveness of the algorithm from both quantitative and qualitative aspects.2)An algorithm based on semi-supervised learning for low-light image enhancement is proposed.Although the current low-light image enhancement algorithms based on deep learning have achieved excellent performance,in the training process,most algorithms still require a large number of paired images for training.Due to the difficulty and cost of obtaining a large number of real low-light/normal images,the use of synthetic images tends to allow the network to learn specific patterns,which makes it not have good generalization ability on real test data.Therefore,a low-light image enhancement algorithm based on semi-supervised learning is proposed,and the synthesized low-light/normal image pairs and the real images are fed into the network learning together.First,the synthetic data is trained in the network,and Gaussian process is used to model the synthetic data in the latent space.Then the real data is fed into the network,and pseudo labels are generated for the real data by using the Gaussian process previously modeled in the synthetic data training stage,which are used for the network training of the real data.The algorithm was tested on a synthetic dataset and five widely used real datasets.At the same time,it was compared with other enhanced algorithms.The experimental results proved the performance of the proposed algorithm.