Fast image quality assessment via supervised iterative quantization method

Abstract

No-reference/Blind image quality assessment (NR-IQA/BIQA) is significant for image processing and yet very challenging, especially for real-time application and big image data processing. Traditional NR-IQA metrics usually train complex models such as support vector machine, neural network, and probability graph model, which result in long computational time and poor robustness. To overcome these weaknesses, the paper proposes a fast no-reference image quality assessment via hash coding method, named NRHC. First, the image is divided into overlapped patches to extract the spatial statistical features of natural scene images. Then the features are encoded to produce binary hash codes via supervised iterative quantization (SITQ) method. Finally, the Hamming distances between the hash code of the test image and those of original undistorted images are calculated to obtain the final image quality. Thorough experiments on benchmark databases demonstrate that the proposed approach achieves comparable performance and has higher computational efficiency and stronger robustness compared with the state-of-the-art NR-IQA methods.

Introduction

With the tremendous development of intelligent network, ultra-high resolution display, and wearable devices, high quality and credible visual information (e.g., image and video) is significant for the end user to obtain a satisfactory quality of experience (QoE). Assessing the quality of visual information, especially no-reference or blind image quality assessment (NR-IQA or BIQA) method, plays an important role in numerous visual information processing system and applications [1]. Moreover, effective (high prediction accuracy) and efficient (low computational complexity) NR-IQA method is essential and has attracted a large number of attentions.

NR-IQA metric is designed to automatically and accurately predict image quality without reference images. It is a difficult and challenging work, but it has attracted many researchers' attentions. Traditional methods focus on designing distortion-specific methods [2], [3], [4], which means that these methods evaluate quality of images with only one kind of distortions effectively, such as JPEG or JPEG2000 compression distortions, white noise, and Gaussian blurring. Therefore, it is imperative to build the general purpose NR-IQA metric to handle different types of distortions and even multi-distortions.

Recently, great effort has been made to develop general purpose NR-IQA metrics named distortion-agnostic NR-IQA method, which can predict quality of images without known the type of distortions. Almost all of the proposed NR-IQA methods [5], [6] contain quality-aware feature extraction and effective evaluation model designing, which are two key processes for building a distortion-agnostic metric. Generally, natural scene statistical (NSS) properties [7] are most popular utilized features, which are extracted by generalized Gaussian distribution (GGD) in wavelet domain usually. Also other features are extracted through Gabor in spatial domain [8] or statistical characteristics in discrete cosine transformation (DCT) domain [9]. All of features which are extracted using statistical method can reveal the naturalness of natural scene images. Another key point is designing the prediction modeling, which can be divided into two categories, two-steps strategy and transductive approach. The former first determines the type of distortions in the test image and then employs an associated distortion-specific no-reference image quality assessment metric to predict the quality of the given image, e.g. BIQI [5] and DIIVINE [10]. The BIQI trains a support vector machine (SVM) model to divide five different types of distortions and trains five different support vector regressions (SVR) model for a particular distortion to predict image quality. The DIIVINE, which is the extended work of BIQI, also is built in the two-steps framework. While the transductive approach aims to build a model to directly project image features to image quality without distinguishing different types of distortions, such as LBIQ [11], BLIINDS [12], BLIINDS-II [9], CORNIA [8], [13], NIQE [14], and SRNSS [15]. In those metrics, a large number of machine learning methods are utilized to train the quality prediction model, such as multiple kernels learning (MKL) [16], neural network [17], [18], and the probabilistic model [19]. Therefore, the reported metrics face a significant problem that they need long training and test time. This is because that complex machine learning model is adopted, the parameters are mostly defined by experience, and a large number of samples are utilized to train the prediction model. And these methods also would reduce the robustness of the quality evaluation system.

In order to solve the above problems, this paper proposes a novel no-reference image quality assessment metric via supervised iterative quantization method, which is simple yet very fast. The proposed method first divides the image into overlapped patches and extracts the image spatial quality evaluator features for each patch. Then the features are encoded into hash codes via a supervised iterative quantization method. Finally, the Hamming distance between the hash code of the test image and the original undistorted image is calculated to predict image quality. In the proposed method, the quality prediction includes hash coding and the Hamming distance calculation [20], [21]. They have the properties of fast speed and high efficiency. Hence, the proposed can satisfy the real-time applications and big image data processing.

The rest of the paper is organized as follows. Section 2 illustrates the proposed no-reference image quality assessment method. Detailed experimental results are summarized and discussed in Section 3, and Section 4 concludes the paper.