Combination of Deep Learning-Based and Handcrafted Features for Blind Image Quality Assessment

This last decade, a plethora of handcrafted-based image quality metrics has been proposed in the literature. Some of them are based on structural analysis, while some others exploit mutual information or perceptual characteristics. Nowadays, deep learning-based methods are widely used in several domains due to its ability to well fit the target directly from the image. In this paper, we study the impact on the performance of combining handcrafted and Deep Learning-based (DL) features, since each of them extracts specific information. Indeed, DL-based image quality assessment methods often extract local information by extracting small patches, while the handcrafted ones provide global information through a global analysis. We analyzed the performance before and after combining the two using bilinear pooling strategy. Experimental results on commonly used datasets show the relevance of combining both approaches.