Research And Implementation Of The Algorithm For Fine-grained Object Classification

Over the last few years, computer version has become an increasingly hot research area. In order to achieve effective management, organization and reuse of image data, image semantic analysis and understanding are developed into extreme active research direction. In the field of computer vision, image cognition are generally divided into three levels. The first level is perception level based image content analysis, it mainly focuses on low-level characteristics extraction and processing, for example, texture, color and temporal relations, etc. The second level is cognitive layer based image understanding, it mainly concentrate on middle characteristics extraction and semantic understanding, including images, scenes, the main area of the video, etc. The third level is emotional layer based image sentiment analysis, which mainly analyzes high-level characteristics, such as facial expression classification and image emotional classification. As the image understanding and expression of human is closer to the cognitive layer, the main scene and object in images are applied for retrieval and classification of objects. This is a core issue in the field of computer vision, which attracts numbers of scholars’attention and research in the recent years.The current image classification are mostly applicable to coarse-grained object categories, for example, the classification between vehicles, buildings and flowers. It lacks the analysis between similar categories, cannot be adopted to achieve further finer distinction and multi-level classification. Therefore, it has great theoretical significance and practical value to develop a new object classification which can provide higher fineness, accuracy and effectiveness.This paper constructs a fine-grained object classification based on vehicle image database. We analyze and verify the effectiveness of two main object classification through experiments.The first method is fine-grained object classification method based on component model. It first implements the model training for objects which is trained by weak mark and latent SVM discrimination iterative algorithm, and then the areas of the object components and middle level characteristics of images are determined by matching pyramid feature and part filters of object model. At last it combines the low level and middle level characteristics to obtain the new image characteristics.The second method is based on convolutional neural networks. It adopts the architecture of 8 layers convolution neural network model. The architecture through the partial normalization response of adjacent nodes in the same layer, overlapping pool in convolution layers and reduce overfitting to optimize the network structure. Then it uses nonlinear rectified linear units as output function. In the end, it uses sixth hidden feature map of convolution neural network as image characteristics to study.After the verification, the two fine-grained object classification method used in this paper are both proved to be effective, especially the convolutional neural network method, which achieved the better classification performance.