Experimental Evaluation of Saliency Detection Model for Product Region Prediction

This paper aims to clarify effectiveness of saliency detection models, which can identify regions of human eye fixation in images, in product region prediction towards further improvement of product image extraction in the future. To confirm the effectiveness, we tested seven handcrafted feature based saliency detection models, which are developed by expert priori knowledge, and three Deep Neural Network (DNN) based saliency detection models, which can perform the automatic feature extraction without much human intervention. Our evaluation, which employed 300 product images, demonstrated that GMR model, which is classified as the handcrafted feature based saliency detection model and which is based on the assumption that the regions in image boundary are more likely to be the background, was the most effective, because prediction accuracy of GMR model was 94.80%, which was 3.47% or more higher than those of other models. Thus, GMR model, which does not require any annotated training dataset, is expected to construct the lower introduction cost and more accurate product image extraction.