Image-Quality Evaluation based on Regional Saliency Pooling

Abstract

Inspired by the successful application of visual saliency in image quality evaluation, we propose an image quality metric based on regional saliency pooling. We first introduce the image saliency detection model to obtain regional saliency maps for image sub-patches. Then, the importance of each image sub-patch is calculated using a VGG16 network based on its saliency map. Such importance is referred to as the quality weight, which is also the pooling result in the proposed framework. Finally, the prediction quality is defined as the weighted sum of the Structural SIMilarity (SSIM) of all image sub-patches. In the experimental part, we choose the popular LIVE image quality database. The results show that the performance of our model is highly competitive, which indicates the effectiveness of the proposed saliency pooling strategy.