K-means Segmentation Based Flower Image Retrieval

The Content-based Image Retrieval (CBIR) technology has been used in several sectors. The goal of a CBIR algorithm is to retrieve semantically similar images in response to a query image. One of the important sectors of CBIR is the flower image retrieval method. Flower image retrieval is a significant and challenging problem in content-based image retrieval. We propose a content-based Segmentation method for retrieving flower images of specified species using the focused part of the flower image from a large database of flower images of various species. In this project work firstly, we preprocessed the flower images using a k-means algorithm which help us segment the main part of the flower image. Secondly, we extracted features of the segmented flower images on basis of three parts: one is the color of the flower, the second is the texture and the third is the shape of the flower. We use HSV, color moments, and auto correlogram features to extract color feature; wavelet moment, Gabor wavelet, and Local Binary Pattern (LBP) features are used to extract texture feature and HOG feature is used to extract shape feature. Those features are independent to characterize all flower images in the database. For a better experiment, we compute our retrieval results for each feature separately like for only HSV, for only color moments, for only HOG, etc. Secondly, we compute results for pairwise features and find comparisons among them. At last, we compute results (for retrieving similar types of flowers images) for a combination of color, texture, and shape features of a query image. The retrieval is accomplished by calculating the similarities between the query image and the database images by employing a set of distance functions. We construct a dataset with 300 flower images of 15 species and test our method on this dataset. When evaluating the results, we noticed that the color feature is the best feature among the other features of an image. We also observed that there was a variation in results for the different distance functions. Some distance functions are giving higher results. Overall the cosine and correlation functions are good for all the features. We also compared our result with a previous method and we found that our method is giving more than 45% more accurate results in retrieving flower images than the previous method.

• Supervision