Search

Recently, deep learning has made impressive achievements in low-light image enhancement. Most existing deep learning-based methods use convolutional neural networks (CNN) by stacking network depth and modifying network architecture to improve feature extraction capabilities and restore degraded images. However, these methods have obvious defects. Although CNN excels at extracting local features, its small receptive field is unable to capture the global brightness, leading to overexposure. The Transformer model from natural language processing has recently produced positive outcomes in a variety of computer vision issues thanks to its excellent global modeling capabilities. However, its complex modeling method makes it difficult to capture local details and takes up many computing resources, making it challenging to apply to the enhancement of low-light images, especially high-resolution images. Based on deep convolution and Transformer characteristics, this paper proposes a Transformer-CNN Parallel Cooperative Network (TCPCNet), which supplements image details and local brightness while ensuring global brightness control. We also changed the calculation method of the traditional Transformer to be applied to enhance high-resolution low-light images without affecting performance. Extensive experiments on public datasets show that the proposed TCPCNet achieves comparable performance against the state-of-the-art approaches. [ABSTRACT FROM AUTHOR]

Deep modal can provide supplementary features for RGB images, which deeply improves the performance of salient object detection (SOD). However, depth images are disturbed by external factors during the acquisition process, resulting in low-quality acquisitions. Moreover, there are differences between the RGB and depth modals, so simply fusing the two modals cannot fully complement the depth information into the RGB modal. To enhance the quality of the depth image and integrate the cross-modal information effectively, we propose a depth enhanced cross-modal cascaded network (DCCNet) for RGB-D SOD. The entire cascaded network includes a depth cascaded branch, a RGB cascaded branch and a cross-modal fusion strategy. In the depth cascaded branch, we design a depth preprocessing algorithm to enhance the quality of the depth image. And in the process of depth feature extraction, we adopt four cascaded cross-modal guided modules to guide the RGB feature extraction process. In the RGB cascaded branch, we design five cascaded residual adaptive selection modules to output the RGB image feature extraction in each stage. In the cross-modal fusion strategy, a cross-modal channel-wise refinement is adopted to fuse the top-level features of the different modal feature branches. Finally, the multiscale loss is adopted to optimize the network training. Experimental results on six common RGB-D SOD datasets show that the performance of the proposed DCCNet is comparable to that of the state-of-the-art RGB-D SOD methods. [ABSTRACT FROM AUTHOR]

Nowadays the problem of image quality assessment (IQA) for screen content images (SCIs) has become a research hotspot as they are ubiquitous in multimedia applications. Although the quality assessment of natural images (NIs) has been continuously developed in the past few decades, few NI-oriented IQA methods can be directly applied on SCIs due to different visual characteristics between them. In this paper, we present a no-reference quality prediction approach considering the content information of SCIs, which is based on dual-channel multi-task convolutional neural network. First, we segment a SCI into small patches and classify them as the textual patches and the pictorial patches. Then, we devise a novel dual-channel convolutional neural network (CNN) to predict the quality of textual patches and pictorial patches. Finally, we propose an effective adaptive weighting strategy for quality score aggregation. The proposed CNN is built on an end-to-end multi-task learning framework, which assists the SCI quality prediction task through the histogram of oriented gradient (HOG) feature prediction task to learn a better mapping between the input patch and its quality score. The adaptive weighting strategy further improves the representation ability of each SCI patch. Experimental results on two largest SCI-oriented databases demonstrate that the proposed method outperforms most of the state-of-the-art no-reference IQA methods and the full-reference IQA methods. [ABSTRACT FROM AUTHOR]

[Display omitted] • The encoder-decoder-based network with multi-task learning is proposed. • Three self-supervised reconstruction tasks are established. • A low-quality fundus image is decomposed into three feature layers in the enhancement process. • An effective fusion strategy is designed based on the features of different layers. • Experiments prove that our method is superior in visual comparison and objective evaluation. Fundus images are used to assist the diagnoses of ocular diseases, and a high-quality fundus image with more details makes clinical diagnostic results more reliable. However, the quality of fundus images is often unsatisfactory due to the turbidity of refractive medium and the doctor-patient cooperation. To enhance the low-quality fundus images, a transformer-based self-supervised network is proposed. During the training phase, an encoder-decoder-based network is introduced. To counteract the drawbacks of establishing long-term dependencies in the convolutional neural network (CNN), the encoder composed of vision transformer and CNN is proposed so that the global and local information of fundus images is fully extracted. On this basis, three reconstruction tasks with self-supervised constraints are designed to enable the network to extract features from different degenerated images. During the testing phase, a low-quality fundus image is decomposed into three feature layers of reverse, illumination, and detail, and then the multi-layer features are fused via the network. To demonstrate the effectiveness of the proposed method, the non-uniform illumination and blurry fundus images are tested. The average scores of NIQE on underexposed, blurred, and overexposed fundus images are 3.03, 2.98, and 2.80, respectively. The average scores of BRISQUE on underexposed, blurred, and overexposed fundus images are 40.32, 40.55, and 39.76, respectively. The average score of subjective evaluation by three ophthalmologists is 61.17%. Compared with the existing methods, the proposed method achieves the superior performance in both subjective and objective evaluations. [ABSTRACT FROM AUTHOR]

Recently, blind image quality assessment (BIQA) has been intensively studied with deep learning. However, the limited quality-annotated datasets restrict its further development. Although patch-based methods have been leveraged to generate more training data, they usually assign the image quality score to all patches in an image. Consequently, much noise would be introduced. To avoid this, we propose a method learning distance distribution from quality levels to expand the training data, which is pseudo Siamese network-based no-reference image quality assessment (NR-IQA). Specifically, we firstly use the K-means clustering algorithm to classify the quality scores of distorted images into five levels. Subsequently, the pseudo Siamese network is adopted to learn the distance distribution from these quality levels. We mainly decrease distances between the quality scores of images from the same quality level and increase distances between the quality scores of images from different quality levels, in which the distorted images can be compared across different distortion types. Finally, we introduce a fusion layer to average the quality scores learnt by the two branches of the trained pseudo Siamese network, and take fine-tuning on this basis to learn image quality score from one image. To demonstrate the effectiveness of the proposed approach, we evaluate it on benchmark databases (LIVE, TID2013 and LIVE MD). The result shows superior performance to some widely used NR-IQA methods and full-reference IQA (FR-IQA) methods. Cross-database evaluation verifies high generalization ability and high effectiveness of our model. • We propose NR-IQA by learning distance distribution from quality levels. • The pseudo Siamese network is employed to learn the distance distribution. • A fusion layer is proposed to average the quality scores learnt by the Siamese network. • Cross-database evaluation verifies high generalization ability and effectiveness of our model. [ABSTRACT FROM AUTHOR]