Research And Application Of Image Recognition Based On Convolutional Neural Network

With the rapid development of computer technology,deep learning has made breakthrough progress in many fields.And the improvement of hardware conditions has gradually reduced the training time of deep network.So,deep learning has become a hot topic in recent years.Convolutional Neural Network(CNN)is a typical algorithm in deep learning.It has been widely used in the field of image processing and achieved good results by relying on its own characteristics such as local perception field,weight sharing,pooling and sparse connection.In this paper,the method of image recognition based on deep convolutional neural network was studied.For the recognition of single target image and multi-target image.The residual network(ResNet)and YOLOv2 were studied respectively.The specific work is as follows:1.In order to study the influence of different shortcut connection on the accuracy in ResNet.Different residual networks were constructed in this paper according to different shortcut connections.Starting from different depths of the network,a lot of experiments were done with cifar-10 dataset to compare the performance about these networks.The experimental results show that smaller shortcut connection has higher recognition accuracy.The 110-layer Shortcut3-ResNet constructed in this paper can achieve 93.76% recognition accuracy.It is equivalent to the recognition rate of the original 152-layer ResNet.2.In view of the poor recognition of weak targets by YOLOv2 algorithm.The original YOLOv2 was improved in this paper.Firstly,in order to extract the fine-grained features of images,the size of feature map at the last few layers was changed from 13×13 to 26×26.Secondly,in order to enhance the nonlinear capability of the network,a 1×1 convolutional layer was added.In addition,the impact on loss function is different according to the target size.So,the loss function was also improved.Finally,the improved model is verified on PASCAL VOC2007 dataset.The experimental results show that the improved YOLOv2 can more accurately identify the target and improve the poor recognition of small targets.The mAP value is from the original 72.73% improved to 73.20%.3.Infrared images of three model cars were collected using FLUKE infrared thermal imager.A small infrared database was constructed and made into the VOC format.The performance of the improved YOLOv2 was tested on the infrared data.