Optimizing image segmentation by selective fusion of histogram based K-means clustering

We present a simple, reduced-complexity and efficient image segmentation and fusion approach. It optimizes the segmentation process of coloured images by fusion of histogram based K-means clusters in various colour spaces. The initial segmentation maps are produced by taking a local histogram of each pixel and allocating it to a bin in the re-quantized colour space. The pixels in the re-quantized colour spaces are clustered into classes using the K-means (Euclidean Distance) technique. The initial segmentation maps from the six colour spaces are then fused together by various techniques and performance metrics are evaluated. A selective fusion procedure is followed to reduce the computational complexity and achieve a better segmented image. The parameters considered for selection of initial segmentation maps include entropy, standard deviation and spatial frequency etc. The performance of the proposed method is analysed by applying on various images from Berkeley image database. The results indicate an increased entropy in the segmented image as compared to other methods along with reduced complexity, processing time and hardware resources required for real time implementation.