Content based image retrieval for leaf identification using structural features and neural networks

With the advances in digital imaging techniques, the number of digital images has risen in the past decade. With the increasing number of image databases consisting of a large number of images, it is challenging to extract images specific to our requirement. Therefore, it becomes necessary to design efficient search systems to query large image databases. Content-based Image Retrieval (CBIR) algorithms analyze the contents of an image rather than the metadata associated with it. This allows image queries to be processed in a way that extracts images from the database, having a close similarity with the query image. Conventional CBIR systems using cosine similarity have a problem of associating images having maximum content overlap, thus similar orientation and size. Therefore, images relevant to the query but having different sizes and orientation are given lower priority in their results. In this article, we propose an approach which solves this problem and attempts to return the results relevant to the query, irrespective of the contents' orientation and size, using appropriate feature extraction techniques. 32 different types of leaves form a preliminary database of 1427 images. The system performance is assessed from the query results of 16 new images belonging to each of the 32 leaf classes. Features such as leaf shape, structure, color, and texture are extracted for pattern matching. Neural network using radial basis function is used for leaf classification to aid query result retrieval. The base accuracy of the CBIR system is 96.88%. Using a simple human relevance feedback mechanism, the neural network trains itself to capture trends in existing leaf classes. As a result, the accuracy of the CBIR system increases as the database is queried. Features extracted from correctly classified leaves are added to the training set. Thus, patterns in leaves belonging to similar and dissimilar classes are captured by the neural network. This increases the robustness of the CBIR to establish maximum possible precision and recall rates along with accurate retrieval of orientation and size varying results.