Design And Implementation For Fine-Grained Object Classification

Image understanding in semantics has become a more and more popular research area in computer vision. In modern computer vision research, understanding of image content has been divided into three levels:Content analyzing based on perception level (low-level feature), such as color, texture, shape, contour, motion, etc. Image understanding based on cognitive level (middle-level feature), such as concept and semantic information detection for main areas of image as well as objects and scenes, etc. Image analyzing on emotion level (high-level feature), such as emotion classification, expression classification and analysis of aesthetics, etc. Perceptual level features are the data information which can be obtained by computer from images directly, but for the problems like complexity, instability and diversity, the relationship with image semantics can’t build well from the low-level features only. The comprehension and description of images are more close to the analysis on cognition. The image retrieval is usually based on the classification for main objects and sense extracted from images. Hence, the research on the classification of objects as the core issues attracted more researchers in recent years.The existing classification of image objects are based on the classification of coarse-grain, just like the difference analysis among animals, birds and automobiles, the consideration of nuances between content similar images is not enough, it is also difficult for further distinguishing of images with certain category, so multilevel classification requirements cannot be guaranteed perfectly. To classify images with fine, accurate and effective properties which keeps abreast with the needs of users and the contents of images, and generates classified information which is convenient for comprehension and operations by users, is of great theoretical significance and value for applications.This paper takes fine-grained object classification as a starting point, further studies two main methods of object classification, constructs a large-scale fine-grained image database, and is validated and analyzed by experiments in the meantime.1. Fine-grained object classification based on the deformable part-model. Firstly, object model is trained by weak mark and latent SVM discrimination iterative algorithm. Second, part-region of object and mid-level feature of image are identified by matching part filters of object model and pyramid feature of the image. Then, image features are obtained by fusing mid-level features and low-level features. Finally, experiments and analysis are done on our fine-grained animal category classification database.2. Fine-grained object classification based on the convolution neural networks (CNNs). This method utilizes the architecture of8layers convolution neural network model. This architecture uses several methods which use nonlinear rectified linear units as output function, local normalize response of adjacent nodes in the same layer, make convolution layers overlapping pooling, reduce overfitting of fully connected layers to optimize network structure. Finally, the feature map of the6th hidden layer of the convolution neural network is made as image feature. Experiments and analysis are conducted on a fine-grained animal category classification database.These two fine-grained object classification methods achieve better performance. Method based on the CNNs is better than method based on the part-model.