Generative Adversarial Network Based Visual Saliency Prediction with Cascaded Hierarchical Atrous Spatial Pyramid Pooling

Abstract

Visual saliency refers to an area of an image that attracts human attention. The Human Visual System (HVS) can focus on specific parts of a scene, rather than the whole image. Visual attention describes a set of cognitive procedures that choose important information and filter out unnecessary information from cluttered visual scenes. Images become a soul in computer vision since it contains plenty of information and human being receives 80% of information through vision. In processing the whole image while only a certain part of an image is needed, more resources are consumed. Instead of processing the whole pixels of an image, specifying only the needed pixel is computationally efficient to minimize the efforts. This is achieved by using GAN with CHASPP module and EfficientNet-B7 which uniformly scales up all dimensions of the image (depth, width, and resolution) is selected as feature extractor in this study which improves the way of extracting features in visual saliency prediction. Different datasets like CAT2000, MIT1003, DUTOMRON, and PASCALS are used in this study to illustrate the efficiency of the selected models and techniques. Human attention modeling focuses on a bottom-up approach that computes the impact of visual stimuli popping from its surrounding. However different models and algorithms have different results in the prediction of the attention area of an image. In this study, we developed effective visual saliency prediction using GAN with CHASPP and other factors like edge loss and perceptual loss. CHASPP module scored the best result on the same datasets measured by different evaluation metrics. It improved the baseline work of SalGAN+ASPP from 3.356 ± 0.04 to 3.851 ± 0.01 (SalGAN+CHASPP+e). This study concluded that the CHASPP module, edge loss, and perceptual loss have a great influence on visual saliency prediction using a generative model.