Candidate
Han Yin
Title
Multipurpose Image Quality Assessment for Both Human and Computer Vision System via Convolutional Neural Network
Supervisor
Kshirasagar Naik and Akshaya Mishra (Systems Design)
Abstract
Computer vision algorithms have been widely used for many applications, including traffic monitoring, autonomous driving, robot path planning and navigation, object detection and medical image analysis, etc. The performance of computer vision algorithms are highly correlated with the quality of input signal. Images and videos are typical input to computer vision algorithms. The quality of videos and images are impacted by vision sensors; environmental conditions, such as lighting, rain, fog and wind. Therefore, it is a very active research issue to determine the failure mode of computer vision by automatically measuring the quality of images and videos.
In the literature, many algorithms have been proposed to measure imaged video qualities using reference images. However, measuring the quality of image and video without using an reference image, known as no-reference image quality assessment, is a very challenging problem. Most existing methods use a manual feature extraction and a classification technique to model image and video quality. Internal image statics are considered as feature vectors and classical machine learning techniques such as support vector machine and naive Bayes as the classifier. Using convolutional neural network(CNN) to learn the internal statistic of distorted images is a newly developed but efficient way to solve the problem. However, there are also new challenges in image quality assessment field. One of them is the wide spread of computer vision systems. Those systems, like human viewers, also demand a certain method to measure the quality of input images, but with their own standards. Inspired by the challenge, in this thesis, we propose to build an image quality assessment system based on convolutional neural network that can work for both human and computer vision system. In specific, we build 2 models: DAQ1 and DAQ2 with different design concept and evaluate their performance. Both model can work well with human visual system and outperform most former state-of-art IQA methods. On computer vision system side, the models also show certain level of prediction power and reveal the potential of CNNs in facing this challenge. The performance in estimating image quality is first evaluated using 2 standard data-sets and against three state-of-the art image quality methods. Further, the performance in automatically detecting the failure mode computer vision algorithm is evaluated using Miovision's computer vision algorithm and datasets.