A generalized Neuro-Fuzzy Based Image Retrieval system with modified colour coherence vector and Texture element patterns

Abstract

A generalized Neuro-Fuzzy Content Based Image Retrieval (CBIR) system is proposed. The system is trained using General Reflex Fuzzy Min-Max Neural Network (GRFMN) where it can take any number and type of different input features. The existing architecture is simplified and the system is trained in pure clustering mode for colour and texture features. The concept of class labels assigned for each hyperbox after unsupervised training adds the flexibility and robustness. By controlling user defined parameters, the system can categorize images as per the users need. With modifications in bucket size and connecting components better coherency is added to understand the colour contents of the image. Further with selected texture element patterns derived from Texture Unit Spectrum (TUS), better feature vector is obtained for training the GRFMN. With this improved feature extraction combining heterogeneous features, the proposed CBIR system gives an efficient automated retrieval of similar images.