Research On Low Light Image Enhancement Network Based On Multi-Scale Feature Extraction

Due to the influence of extreme lighting environment or the limitation of hardware technology,such as inadequate or unbalanced lighting conditions,extreme backlight conditions and under-exposure during shooting,the shot image will be too dark or too bright,affecting people’s visual observation and unable to convey complete information.In addition,it will also affect the performance of some high-level computer vision tasks,including object recognition,scene classification,etc.For example,in the field of traffic safety,due to insufficient brightness or uneven illumination of street lamps,the visibility at night is very poor,and a large amount of information is lost in the images and videos obtained by road monitoring equipment,which seriously affects the system judgment.Therefore,low light image enhancement technology has great demand and prospect in many fields.Although many traditional methods can cope with the problem of insufficient image brightness and difficult-to-obtain information,such as histogram equalization,gamma correction,and Retinex decomposition,these methods ignore the noise problem caused by uneven lighting,and the enhanced images often appear blurry and with artifacts.In recent years,deep learning methods have made significant progress compared to traditional low-light image enhancement algorithms in terms of robustness and generalization,but there are still the following problems:(1)Unable to effectively restore the extremely dark area of the image;(2)cannot accurately solve the image color deviation problem;(3)Image artifacts brought by image compression cannot be fully recovered.In order to solve the problems of noise and color deviation in this field,this thesis adopts the theoretical method of deep learning and proposes two low light image enhancement methods.The following are the main innovations of this thesis.(1)To solve the problem of detail loss and noise amplification caused by the enhancement method based on traditional Retinex model in the process of calculating the reflectance component of images,this thesis proposes a multi-scale attentional low light image enhancement network based on Retinex theory,which combines deep convolutional neural network with Retinex theory.Learn a more realistic reflectivity layer.First,the low light image was decomposed by Retinex theory to obtain the light focus map,which was fed into the network model together with the low light image for feature extraction.Then,a multi-scale feature enhancement module and a feature fusion module are constructed to realize the learning of the backlight image.Finally,the Retinex model was used to solve the final enhancement results.Experimental results show that this network can obtain enhanced images with good subjective feelings and objective indicators.(2)Considering the color deviation and image texture loss caused by feature information loss in some existing deep learning methods in the process of feature learning,this thesis proposes a low-light image enhancement algorithm based on multi-scale feature fusion.In this method,the low light image is firstly transformed by nonlinear transformation,and the preliminary enhanced image is obtained.Then,the initial enhanced image and the low brightness image are simultaneously fed into a multi-scale feature extraction block for feature extraction.In order to reduce the depth of the network and increase the network’s learning of semantic features,a U-shaped feature enhancement module is constructed in the feature extraction branch of each scale to increase the feature extraction of context information of each scale.Finally,the feature information extracted from multiple scale extraction branches is combined and integrated to obtain the enhanced image.The experimental results on a large number of data sets show that the proposed method achieves better visual effects and better index enhancement results,and is superior to most mainstream low-light image enhancement algorithms.