Performance analysis of saliency detection in images

Abstract

Detecting salient objects in an image has a vast number of applications across the web, and mostly it is done manually. Since the internet is growing faster than ever, it is not a feasible solution for high-scale dynamic applications. Hence, image processing automation with computer vision and machine learning has become a burning topic of research in the past few years. Saliency based image cropping is a task of identifying the notable segments in an image and be able to accommodate these parts in the crop ratio for any given specific viewport. It helps not only in image cropping but also with object recognition, visual tracking and visual restoration, etc. To address this, a novel approach for saliency detection is proposed based on heat map analysis and boundary prior. In this approach, the images go through heat map analysis to identify salient objects based on machine learning model. After that, the images are segmented based on color and texture difference and then major contours are identified. The contours and the obtained saliency coordinates are accommodated inside the crop for each viewport requirement. The crop coordinates then go through two more processes, firstly a shift center process where the crop center is moved towards the important but lesser salient object, and then an inclusivity rule checks that the image is not cropped at coordinates without any salient objects. The simulation results reveal that the proposed algorithm attains better results than the cutting-edge algorithm of Twitter with similarity index of 89.89%.