Research On Low Illumination Image Enhancement And Classification Based On Convolutional Neural Network

As an important research topic,image classification mainly includes image classification methods based on traditional manual feature extraction,semantic separation and convolutional neural network.Among them,the convolutional neural network has received widespread attention because of its ability to extract image features through self-learning,and its strong ability to recognize and classify images.However,the learning process of convolutional neural networks requires a large amount of data samples,and in most cases it is difficult to obtain a large number of high-quality image data sets.To solve this contradiction problem,we need to dig out the useful information in a small amount of low-quality image samples as much as possible,and maximize the classification effect of the model.In this thesis,aiming at a class of low-illuminance image samples,we will propose a low-illuminance image classification method under small samples.The main innovations are as follows:(1)A low-light image enhancement method combining detail compensation network and traditional methods is proposed.First,the image is divided into low-frequency components that characterize the overall brightness and high-frequency components that characterize local details through a filtering algorithm;secondly,the detail compensation network is used to learn additional detail information from multiple image components,and the trained network model is used to control low illumination the details of the image are supplemented;again,the brightness of the low-frequency component information of the image is improved based on the LIME algorithm;finally the enhanced image is fused,and the distortion phenomenon that appears is adjusted locally by network training.(2)An image expansion method based on derivative transformation is proposed.Firstly,the original image is derivatized and transformed based on the derivation function.Derived images with different brightness and contrast are obtained according to the differences produced by different derivation transformations,which increases the number of pictures in the data set and expands the differences between similar images;secondly,based on several differences The expansion method of the image data is further mixed and expanded;finally,the random perturbation method is used to enhance the antagonism during the model training process,thereby enhancing the generalization ability of the model.(3)A small sample low-illuminance image classification method is proposed.First,construct a low-illuminance small-sample data set,decompose the selected small-sample images with high frequency and low frequency,perform gamma transformation and linear transformation on the decomposed low-frequency components,and merge them with the high-frequency components to complete the low-light treatment,and then pass Add random disturbances to further increase the difference of the images in the data set;secondly,combine the contents of the first two chapters to enhance and increase the image quality and capacity of the constructed small sample low-illuminance image classification data set;finally,load the pre-trained model weight transfer idea Perform generalization training on the network model parameters,and then use the processed data set to perform local reinforcement training on the network model parameters,so as to achieve high-quality classification of low-illuminance images in the case of small samples.