Research On Image Algorithm Based On Convolutional Neural Network

Computer vision refers to the machine vision which uses camera equipment and computer to simulate human vision to identify,track and measure the target,through the recognition and analysis of the image to further process of the graphics and make it more suitable for instrument analysis and for human to observe.Computer vision can perceive 3D scene,recognition and understanding in the objective world,realize visual perception,image recognition,face recognition,target location,etc.Computer vision and related algorithms are of great research value in modern society.The advantage of convolution neural network in local perception and information capture makes it play an important role in the research of computer vision.The convolution neural network can abstract the local features of the image at different scales by using the operation of multiple convolution layers.It has strong network expression ability and mathematical mapping characteristics.The deep neural network can solve the problem of over-fitting caused by the small amount of data or too many feature points.Then with the development of the convolutional neural network in image research,the target recognition has more improvement.However,the problems of data dependence and computing resource increase due to the introduction of more features in the network layer are urgent to be solved.At the same time,how to transfer the trained network features to unknown fields is also worth applying.Therefore,the study of feature migration and network adjustment of convolutional neural networks is of great significance.At present,the mainstream convolution network still relies on large-scale training data.Based on the full study of the convolutional neural network in the target recognition algorithm,this paper proposes four improvements from the perspective of performance improvement:1.In order to solve the problems such as too many characteristic parameters in the upper layer of neural network resulting in data redundancy and large computational complexity the method of using maximization pool instead of mean pool for pool layer and multi-layer full-connection layer are used to adjust the network and reduce the parameters.2.Based on the transfer learning algorithm,the trained target recognizer in horizontal perspective and the kmeans algorithm are used for the clustering transfer of the feature parameters to complete the transfer learning.3.Aiming at the problem that Gao Si of Softmax classifier loses large inter-class distance,Center Loss,is introduced to adjust the inter-class distance.4.Completed the design of GPU parallelization calculation with Tensor Flow framework.The thesis has carried out experiments on VOC and MIT data sets,and the experimental results show that the improved migration algorithm has achieved very competitive results on these data sets,improving the accuracy of the algorithm for target recognition.