Search

There are limitations in existing multi-exposure fused image quality assessment (MEF-IQA) approaches. The first is that the existing approaches are hard to estimate the consistency of luminance distribution in fused images, which can be easily destroyed in fusion process. The second is that the aesthetics estimation of fused images are neglected in existing approaches, which should be considered for the images produced by fusion algorithms to give people an aesthetic experience. In order to solve these problems, a MEF-IQA approach based on the deep network was proposed. First, we adapted a quad path network as luminance error quality evaluator (LE) to extract the features, which can capture the luminance distribution patterns, from fused images. Then, we trained an aesthetic evaluator (AE) to obtain the feature which can describe the aesthetic attributes of fused image. Finally, the two extracted features are combined for predicting the perceptual quality scores. Experimental result clearly shows that the proposed approach can quantify the image quality reliably, and it is highly correlated with subjective quality scores.

The blind image quality assessment (BIQA) metric based on deep neural network (DNN) achieves the best evaluation accuracy at present, and the depth of neural networks plays a crucial role for deep learning-based BIQA metric. However, training a DNN for quality assessment is known to be hard because of the lack of labeled data, and getting quality labels for a large number of images is very time consuming and costly. Therefore, training a deep BIQA metric directly will lead to over-fitting in all likelihood. In order to solve this problem, we introduced a weakly supervised approach for learning a deep BIQA metric. First, we pre-trained a novel encoder-decoder architecture by using the training data with weak quality annotations. The annotation is the error map between the distorted image and its undistorted version, which can roughly describes the distribution of distortion and can be easily acquired for training. Next, we fine-tuned the pre-trained encoder on the quality labeled data set. Moreover, we used the group convolution to reduce the parameters of the proposed metric and further reduce the risk of over-fitting. These training strategies, which reducing the risk of over-fitting, enable us to build a very deep neural network for BIQA to have a better performance. Experimental results showed that the proposed model had the state-of-art performance for various images with different distortion types.