Research On Scene Image Classification And Weak Light Image Enhancement Based On CNN

Images contain a large amount of information,which is a direct and efficient information source for human perception of the world,and an important carrier for human expression of information and an important medium for transmitting information.With the rapid development of hardware and computer network technology,the number of digital images is exponentially increasing.It is an imperative trend to use computers to process a large number of digital images to complete related computer vision tasks.Among many computer vision task solving methods and tools,convolutional neural network(CNN),as an excellent image feature extractor,has been widely used in image classification,target recognition and other fields,and achieved good results.However,because CNN needs a lot of images to support training,its convergence speed and over-fitting restrict its performance In some specific scenarios and specific applications,CNN structure still needs to be optimized to achieve better results.In this paper,based on the existing literature,this paper studies scene image classification and low illumination image enhancement based on convolution neural network.The main research contents include(1)To solve the problem that the existing scene recognition algorithm based on convolution neural network cannot deal with the multi-spectral image of the target scene,a multi-spectral scene recognition method based on multi-way convolution neural network is proposed.Using the improved CNN structure to synthesize multi-spectral information combined with the pre-training method,the classification accuracy of scene is effectively improved(2)In order to reduce the dimension of information,general CNN using pooling layer which results in loss of information and affects the expressive ability of network.To solve the problem,a Parameter Pooling Layer is proposed to replace the pooling layer in general CNN.With only a small number of network parameters added to the parameter pooling layer,it is possible to retain the desired features of convolutional neural network and improve the network performance.At the same time.,the forward propagation information of the parameter pooling layer is added,which affects the weight updating of the back propagation algorithm,and the network is more easily converged and converges faster(3)A low-illumination image enhancement method based on de-illumination effect is proposed.Convolutional neural network is used to learn the parameters needed to be estimated in Retinex theory from the training set of low-illumination images to achieve end to-end low-illumination image enhancement.Compared with the classical algorithm and other CNN-based methods,the simulation data experiments and real low illumination image enhancement experiments show that the algorithm effectively improves the overall brightness and contrast of the image,and improves the structural similarity and peak signal-to-noise ratio(PSNR)of the image.