Research And Implementation Of Object Detection Based On Convolutional Neural Networks

As a direction of computer vision research, Object recognition is currently great hot aerea of reaserch. At present, with the material needs of people increasing gradually, object recognition is playing a pivotal role in the field of security, science, technology and the economy. It has the importance in the demands for security surveillance and traffic system, so the study of object recognition has a very important significance for society in the future. The Appearance of deep neural network which is a new method marks the arrival of a new era for the deep learning. As one of deep learning approach, convolution neural network has excellent performance and great potential, which becomes the reason to be hot research subject and gets success, therefore the study of object recognition based on CNN has become a social hot spot in the field of Object recognition.The thesis focuses on a new design of the convolutional neural network model that its performance has improved significantly. The main works are as follows:1. the improved model modifies the loss function of YOLO network. Improved model based on YOLO replaces margin style with proportion style. Compared to the old loss function, the new is more flexible, and more reasonable to the optimization of the network error.2. The improved model adds convolution kernels of 1 * 1, so as to reduce the number of weight parameters of the layers and to achieve the smooth transition of extracted features.3. The improved model uses inception module structure. Because Inception module has the ability to deepen and widen the network, and to reduce a lot of network parameters.4. The improved model adds a Spatial Pyramid Pooling layer, which can make full use of the information of image and enhance the recognition performance of the network. The overall time performance of the network is also improved because of the advantages of SPP.Finally, the improved convolutional neural network based on CNN experiments were carried out through the pascal voc2007 database and Pascal voc2012 database, using confusion matrix, time complexity and information visualization tools for collecting experimental data to analyze and summarized.The experimental results shows that the new network are better than other networks in recognition performance and in time efficiency,also have a certain competitiveness in the international advanced level of networks.The thesis apply the network model to practical application, and implemented a new system based on YOLO because of the better performance in the time performance.