Recognizable or Not: Towards Image Semantic Quality Assessment for Compression

Traditionally, image compression was optimized for the pixel-wise fidelity or the perceptual quality of the compressed images given a bit-rate budget. But recently, compressed images are more and more utilized for automatic semantic analysis tasks such as recognition and retrieval. For these tasks, we argue that the optimization target of compression is no longer perceptual quality, but the utility of the compressed images in the given automatic semantic analysis task. Accordingly, we propose to evaluate the quality of the compressed images neither at pixel level nor at perceptual level, but at semantic level. In this paper, we make preliminary efforts towards image semantic quality assessment (ISQA), focusing on the task of optical character recognition (OCR) from compressed images. We propose a full-reference ISQA measure by comparing the features extracted from text regions of original and compressed images. We then propose to integrate the ISQA measure into an image compression scheme. Experimental results show that our proposed ISQA measure is much better than PSNR and SSIM in evaluating the semantic quality of compressed images; accordingly, adopting our ISQA measure to optimize compression for OCR leads to significant bit-rate saving compared to using PSNR or SSIM. Moreover, we perform subjective test about text recognition from compressed images, and observe that our ISQA measure has high consistency with subjective recognizability. Our work explores new dimensions in image quality assessment, and demonstrates promising direction to achieve higher compression ratio for specific semantic analysis tasks.