Deep Algorithm Quality (DAQ): A blind computer vision algorithm quality assessment technique
Huanyu Yin, Anima Mishra, Kshirasagar Naik · 2017
No-Reference image quality assessment is a challenging problem of great interest to computer vision research community. This paper proposes to find a general solution to measure image quality for both human and computer vision system. A supervised deep neural network, called Deep Algorithm Quality (DAQ), is designed to blindly measure the human visual quality of benchmark images as well as to predict the performance of a computer vision algorithm on distorted images. The performance of the proposed DAQ is evaluated using two sets of experiments. The first experiment formulates DAQ as an image quality estimator, and evaluates the performance on general image quality assessment benchmarks. DAQ has achieved highest LCC and SROCC scores compare to five state-of-the-art image quality assessment methods. In the second experiment, DAQ is trained to work as a failure detector to predict the detection performance of another predefined computer vision algorithm working on distorted images.