Digital Implementation of Neuro-Fuzzy System for Image Processing Functions

This paper described a hardware implementation approach of a new neuro-fuzzy system (NFS). The main idea was to exploit the powerful means of the adaptive neuro-fuzzy inference system with respect to function approximation, making possible the implementation of reconfigurable hardware with on-chip learning. Different image processing tasks could be achieved based on a back-propagation (BP) learning algorithm. The complexity of this kind of implementation made the pulse mode an attractive solution. Such a technique provided higher integration density through its compactness. Details of the proposed design with on-chip learning were given. As application, illustrating the efficiency and scalability of the proposed NFS, we considered the approximation of image edge detection, which is a very important step in image processing. The proposed system provided efficient learning and good generalization results for different image categories (uniform, synthetic with texture, and natural images). Moreover, the efficiency of our proposed system versus other approaches was demonstrated. Design synthesis results on a Virtex-5 field programmable gate array (FPGA) platform were presented, proving that the implemented NFS provided the best compromise between compactness, speed, and accuracy compared to previous work in the literature.