Multi-task learning approach for natural images' quality assessment

Blind image quality assessment (BIQA) is a method to predict the quality of a natural image without the presence of a reference image. Current BIQA models typically learn their prediction separately for different image distortions, ignoring the relationship between the learning tasks. As a result, a BIQA model may has great prediction performance for natural images affected by one particular type of distortion but is less effective when tested on others. In this paper, we propose to address this limitation by training our BIQA model simultaneously under different distortion conditions using multi-task learning (MTL) technique. Given a set of training images, our Multi-Task Learning based Image Quality assessment (MTL-IQ) model first extracts spatial domain BIQA features. The features are then used as an input to a trace-norm regularisation based MTL framework to learn prediction models for different distortion classes simultaneously. For a test image of a known distortion, MTL-IQ selects a specific trained model to predict the image’s quality score. For a test image of an unknown distortion, MTLIQ first estimates the amount of each distortion present in the image using a support vector classifier. The probability estimates are then used to weigh the image prediction scores from different trained models. The weighted scores are then pooled to obtain the final image quality score. Experimental results on standard image quality assessment (IQA) databases show that MTL-IQ is highly correlated with human perceptual measures of image quality. It also obtained higher prediction performance in both overall and individual distortion cases compared to current BIQA models.