Image Enhancement Technology Based On Multi-scale Feature Extraction Network

Nowadays,the improvement of industrial technology has led to the development of various imaging equipment,and the quality of the images captured is also increasing.However,due to the limitation of objective conditions,it is difficult for the captured images to meet the actual needs.For example,in a low-light environment with a lack of light sources and outdoors at night,due to insufficient light,the images obtained by photographic equipment often have many problems such as loss of details,color deviation,and low sharpness and contrast.However,this kind of environment cannot be avoided,so the post-processing of the captured image through the algorithm has become the only way,and the low-light image brightness enhancement technology is applied.Currently,most image brightness enhancement methods are proposed based on imaging models of low-light images or deep learning techniques.The model-based method will cause color cast and distortion in the enhancement results due to the inaccuracy of the solving parameters;the deep learning-based method can learn the relationship mapping between the degraded image and the clear image well,but it is still highly relying on the reliability of the data set,and the current data set is not complete.Therefore,such methods are the current research hotspot in this field.In this thesis,based on deep learning technology,combined with the characteristics of low-light images,two low-light image brightness enhancement networks are proposed.In addition,in order to make up for the shortcomings of outdoor enhancement datasets in the field of image enhancement today,a new outdoor low-brightness enhancement dataset is constructed.The main innovations of this thesis are as follows:(1)Aiming at the loss of details in the feature extraction process of the current mainstream image brightness enhancement networks,this thesis proposes a multi-scale grid image enhancement network.The network consists of three main modules: a smooth convolution cascade block,a dual-scale feature downsampling fusion block,and a dual-scale feature upsampling fusion block.In order to avoid the loss of details caused by the reduction of the scale of the feature map during the learning process of the network,the network builds a main branch to ensure that the size of the feature map and the input image are consistent,and uses stacking multiple smooth convolution cascade blocks to extract features to reduce the network.parameters and calculations.In order to extract higher-level structural information,the network designs three branches by reducing the size of the original feature map to achieve feature extraction at different scales of the image.At the end of the network,the features of different scales are integrated to obtain the final enhanced image.In order to better train the network,a chromatic aberration loss is introduced in the loss function to ensure more accurate restoration of the color of the image.Extensive experimental results on the public dataset LOL show that the proposed method outperforms existing mainstream image brightness enhancement methods in both objective and subjective results,proving the effectiveness of the method.(2)Most of the existing image brightness enhancement methods only consider the extraction of features on a single scale of the image,which usually leads to the problem of incomplete feature extraction.Therefore,this thesis proposes a multi-scale residual feature fusion image enhancement network based on deep learning.The network consists of two modules: a shallow feature extraction module and a multi-scale feature fusion module.First,a shallow feature extraction module is constructed to extract multi-scale shallow features through simple convolution layers and smooth convolution residual blocks.A smooth convolutional residual block runs through the entire network to extract features and avoid gridding artifacts.Then,a multi-scale feature fusion module is constructed by cascading multiple feature ensemble residual blocks to fuse shallow and deep features of the same scale.Finally,the fused features are enhanced by convolutional layers.Experimental results on the LOL dataset and our proposed outdoor low-luminance enhancement dataset show that the enhancement results of this method can provide more accurate color saturation and richer details compared to some state-of-the-art methods,justifying the method performance and effects.(3)In addition,in order to make up for the defect that most of the current datasets in the low-light enhancement field are synthesized or concentrated indoors,this thesis constructs an outdoor low-light enhancement dataset.Through sufficient experiments,the effectiveness of this dataset is demonstrated.