FEATURE STRUCTURE SIMILARITY INDEX FOR HYBRID HUMAN AND MACHINE VISION

More and more images/videos will be consumed by both human and machine in many fields. Optimization of image processing algorithm for hybrid human and machine becomes a challenging task. To address this problem, feature structure similarity index (FSSIM) is proposed in this paper as an objective metric for image quality assessment (IQA), by defining structure similarity in low-level feature domain. Features extracted by the first convolutional layer of pre-trained resnet50 network are treated as common feature domain for both human and machine vision. Moreover, multi-scale structure similarity with weighting matrix is used as distance measure in the feature domain. FSSIM is capable of fully decoupling image processing and its downstream machine tasks, enabling image processing algorithm optimization for hybrid human and machine vision. Experimental results show FSSIM-optimized image processing algorithms achieve significant performance improvement over existing metrics in context of machine vision tasks including object detection and semantic segmentation. Meanwhile reconstructed images of FSSIM-optimized algorithms are better friendly to human vision.