Research On Fine-Grained Classification For Visual Flower Image

Developments in information technology have led to convenience for people to take flower images. However, the flower image recognition still needs the assistance of professional guidance. The flower image classification belongs to the fine-grained category of image classification. Due to complicated background, inter-class similarity and intra-class difference, traditional image classification method cannot applicable to this kind of flower classification task directly. Considering the above unsolved and challenging issue, in this thesis, we propose an image segmentation and multi-feature fusion method for the automatic flower image classification. The main work content and contribution are as follows:1. We propose a flower image foreground segmentation method based on saliency detection (SDM), in terms of the complicated background. To be specific, we train a foreground and background classifier for flower saliency region, set initial foreground and background segmentation threshold adaptively, and then segment the main part of a flower in the image from the background based on the GrabCut method. Different from the classical flower classification method which need annotation of foreground and background for each category in the training database, we propose a common flower classification method. This method based on saliency detection foreground segmentation and don’t need a separate training for each kind of flower.2. Due to the inter-class similarity and intra-class difference issue, we propose a hierarchical classification method based on multi-feature fusion. The traditional feature fusion methods are only the linear addition of each feature and neglect the influence of different features for different flower categories, while the proposed method in this thesis applies hierarchical classification method based on multi-feature fusion. In this method, we train a classifier using color and shape feature for each type individually and obtain score of each category. Through experiment, this method shows the ability of overcoming the intra-class difference problem and achieves effective segmentation results.The experimental results indicate that, the SDM can distinguish the foreground and background of flower image, and outperforms the state-of-the-art methods which only use the saliency segmentation, In the Oxford flower database, our proposed hierarchical classification method based on multi-feature fusion can classify the flower with high inter-class similarity and intra-class difference.