Flower image retrieval with category attributes

In this paper, we recommend a mid-level representation, category attributes, for content based flower image retrieval. Low-level features have been utilized in images for retrieval. However, even though these features are efficient, the similarity between low-level features may differ from high level human perception, known as semantic gap. In real life, it is very usual to use attributes, a domain specific terminology, to describe the visual appearance of objects. Inspired by this, we utilize category attributes, i.e. daisy, buttercup or iris to construct semantic representation of flower images. For each category attribute, we train a linear SVM based on low level visual features, containing the appearance of color, texture and shape. Outputs of these classifiers are regarded as attribute features. Then we use distances between attribute features for flower retrieval. This method was evaluated on 17 Category Flower Dataset. Experimental results show that attribute based representation outperforms low-level features in terms of mean average precision.