Fine-Grained Visual Categorization With Part Alignment Model

Image classification has been a core problem in machine learning and computer vision. With the continuous development of computer technology and image classification, fine-grained object recognition issue has become a very active research since the general classification cannot meet the retrieval precision people required. The most popular and effective approaches based on local features are the Bag-of-Visual-Words model and the deep Convolutional Neural Network. However, due to the semantic gap between feature representation and attributes as well as the limitations of the concept expression of local visual feature, traditional image representation models often suffer from certain shortcomings, including the insufficient description of detailed area, and the lack of spatial structure information in feature encoding, etc.This paper will apply models based on local features to implement the fine-grained classi-fication. The main works of this thesis are summarized in the following:(1) A comprehensive analysis based on status and challenges of fine-grained classification is proposed in this paper. Fine-grained visual categorization (FGVC) refers to the task of clas-sifying objects that belong to the same basic-level class (e.g., different bird species). However, FGVC remains a major challenge due to the large inner-class variation and subtle inter-class variation. Since the subtle inter-class variation often exists on small meaningful regions which are called parts (e.g., beak), it is reasonable to localize semantic parts of an object before describ-ing it. The method has low computation complexity and does not need many human interaction has been applied to obtain competitive classification accuracy.(2) Based on over-segmentation and ambiguity of corresponding relation problems caused by unsupervised segmentation algorithm, two detailed segmentation algorithms are proposed in this paper. The connectivity and distance of parts are added to the objective function in order to reduce the ambiguity of image representation. Moreover, ultrametric contour map is explored to the training strategy to obtain more reasonable boundaries of parts. The image representations based on fine-tuned parts got better classification accuracies.(3) This paper combines the superpixels and parts-alignment algorithm to solve the fine-grained classification problem. First, graph structure and augmented energy function are con-structed on images. Then a weighted k-means algorithm is used in high dimensional feature space to build superpixels. The class-specific foreground templates are computed by the hi-erarchical trees constructed by superpixels. As the test images share similar appearance with the template images, the resulting segmentations are obtained through image matching. The experimental results demonstrate that the proposed algorithm could provide better image repre-sentation and improve the classification accuracy.The efficiency of the proposed methods is substantiated by a series of experiments using CUB-200-2011 dataset.