Research Of Face Recognition And Multi-focus Image Fusion Algorithms Based On CNN

With the continuous improvement of technology and the continuous improvement of people's life quality,the application of convolutional neural network(CNN)has been emerging in various fields in recent years.CNN is a kind of non-fully connected feedforward neural network,which is a deep learning method that successfully adopts multi-layer hierarchical networks.In this master's thesis,the applications with CNN in face recognition and multi-focus image fusion are studied.A face recognition algorithm and an improved multi-focus image fusion algorithm of CNN are proposed.The main contents and innovations of this master's thesis include:(1)The structure and principle of CNN are introduced in detail.Firstly,from a single neuron to a simple neural network,and then CNN is gradually introduced.The principle of CNN is based on the three basic ideas: local receptive field,shared weight and pooling.That makes the high-dimensional data processing more easily by CNN,then the features of the image are extracted relatively simple and accurate and feature classification effect will be better.(2)The self-normalized convolutional neural network(SCNN)is proposed for face recognition.The algorithm uses SCNN to extract and classify face features.Then the best experimental conditions are found by comparing experimental results with different batch sizes and different network layers.Finally,compared with traditional CNN algorithms and others,the proposed method is tested on the ORL database,and experimental recognition rate is up to 98.3%.Experimental results show that the method based on SCNN has higher recognition rate and faster convergence rate than ordinary convolutional neural networks in face recognition.(3)A new multi-focus image fusion method based on convolution neural network is proposed.The algorithm mainly includes four steps: focus detection,initial segmentation,consistency verification and fusion.Firstly,the two source images are input into a pre-trained CNN model to output a fractional image containing the focus information of the source image,and then the focal image of the source image with the same size is obtained from the fractional image by averaging the overlapping regions,down the use of two popular consistency verification strategy to refine the binary partition diagram to generate the final decision diagram.Finally,a fused image is obtained using the final decision-making graph of pixel-by-pixel weighted averaging strategy.Compared with the other fusion results,the proposed method has obvious advantages both in subjective observation and objective indexes.