We explore image semantic quality assessment (ISQA) for compression of images that are utilized for automatic image analyses, such as recognition and detection, rather than for human viewing. For such analyses purposes, we argue that the quality of compressed images should be evaluated from its preserved semantic-related features, instead of its pixel-wise fidelity (e.g. PSNR) or visual quality (e.g. SSIM). In this paper, we make an empirical study of an ISQA approach based on SIFT features extracted from both original and compressed car-plate images, and we formulate an optimization problem to find the operating point of an image compression system for car-plate recognition. Experimental results show that our proposed ISQA measure is significantly better than PSNR and SSIM in predicting the recognizability of compressed car-plate images. Accordingly, using our ISQA measure during compression leads to more than 50% bit-rate saving compared to using PSNR or SSIM.