In this paper we propose a new no-reference (NR) image quality assessment (IQA) metric using the recently revealed free-energy-based brain theory and classical human visual system (HVS)-inspired features. The features used can be divided into three groups. first involves inspired by free energy principle structural degradation model. Furthermore, also reveals that HVS always tries to infer meaningful part from stimuli. terms of finding, predict an perceives distorted based on theory, then...

We propose a variational Bayesian scheme for pruning convolutional neural networks in channel level. This idea is motivated by the fact that deterministic value based methods are inherently improper and unstable. In nutshell, technique introduced to estimate distribution of newly proposed parameter, called saliency, on this, redundant channels can be removed from model via simple criterion. The advantages two-fold: 1) Our method conducts without desire re-training stage, thus improving...

The general purpose of seeing a picture is to attain information as much possible. With it, we in this paper devise new no-reference/blind metric for image quality assessment (IQA) contrast distortion. For local details, lirst roughly remove predicted regions an since unpredicted remains are information. We then compute entropy particular areas maximum via visual saliency. From global perspective, compare the histogram with uniformly distributed symmetric Kullback-Leibler divergence....

Recent works on domain adaptation reveal the effectiveness of adversarial learning filling discrepancy between source and target domains. However, two common limitations exist in current adversarial-learning-based methods. First, samples from domains alone are not sufficient to ensure domain-invariance at most part latent space. Second, discriminator involved these methods can only judge real or fake with guidance hard label, while it is more reasonable use soft scores evaluate generated...

With the widespread adoption of multidevice communication, such as telecommuting, screen content images (SCIs) have become more closely and frequently related to our daily lives. For SCIs, tasks accurate visual quality assessment, high-efficiency compression, suitable contrast enhancement thus currently attracted increased attention. In particular, evaluation SCIs is important due its good ability for instruction optimization in various processing systems. Hence, this paper, we develop a new...

In this paper, we propose a new psychovisual quality metric of images based on recent developments in brain theory and neuroscience, particularly the free-energy principle. The perception understanding an image is modeled as active inference process, which tries to explain scene using internal generative model. thus closely related how accurately visual sensory data can be explained by model, upper bound discrepancy between signal its best description given free energy cognition process....

In this paper, we investigate the problem of image contrast enhancement. Most existing relevant technologies often suffer from drawback excessive enhancement, thereby introducing noise/artifacts and changing visual attention regions. One frequently used solution is manual parameter tuning, which is, however, impractical for most applications since it labor intensive time consuming. research, find that saliency preservation can help produce appropriately enhanced images, i.e., improved...

Applying the knowledge of an object detector trained on a specific domain directly onto new is risky, as gap between two domains can severely degrade model's performance. Furthermore, since different instances commonly embody distinct modal information in detection scenario, feature alignment source and target hard to be realized. To mitigate these problems, we propose Graph-induced Prototype Alignment (GPA) framework seek for category-level via elaborate prototype representations. In...

In a typical image communication system, the visual signal presented to end users may undergo steps of acquisition, compression and transmission which cause artifacts blurring, quantization noise. However, researches quality assessment (IQA) with multiple distortion types are very limited. this paper, we first introduce new multiply distorted database (MDID2013), is composed 324 images that simultaneously corrupted by JPEG noise injection. We then propose six-step blind metric (SISBLIM) for...

High dynamic range (HDR) imaging techniques have been working constantly, actively, and validly in the fault detection disease diagnosis astronomical medical fields, currently they also gained much more attention from digital image processing computer vision communities. While HDR devices are starting to friendly prices, display still out of reach typical consumers. Due limited availability devices, most cases tone mapping operators (TMOs) used convert images standard low (LDR) for...

To enhance the visibility and usability of images captured in hazy conditions, many image dehazing algorithms (DHAs) have been proposed. With so DHAs, there is a need to evaluate compare these DHAs. Due lack reference haze-free images, DHAs are generally evaluated qualitatively using real images. But it possible perform quantitative evaluation synthetic since available full-reference (FR) quality assessment (IQA) measures can be utilized. In this paper, we follow strategy study DHA...

For monocular 3D pose estimation conditioned on 2D detection, noisy/unreliable input is a key obstacle in this task. Simple structure constraints attempting to tackle problem, e.g., symmetry loss and joint angle limit, could only provide marginal improvements are commonly treated as auxiliary losses previous researches. Thus it still remains challenging about how effectively utilize the power of human prior knowledge for In paper, we propose address above issue systematic view. Firstly, show...

In this work, we propose a Cross-view Contrastive Learning framework for unsupervised 3D skeleton-based action Representation (CrosSCLR), by leveraging multi-view complementary supervision signal. CrosSCLR consists of both single-view contrastive learning (Skeleton-CLR) and cross-view consistent knowledge mining (CVC-KM) modules, integrated in collaborative manner. It is noted that CVC-KM works such way high-confidence positive/negative samples their distributions are exchanged among views...

Crowd counting is a challenging task in the presence of drastic scale variations, clutter background, and severe occlusions, etc. Existing CNN-based methods tackle these challenges mainly by fusing either multi-scale or multi-context features to generate robust representations. In this paper, we propose address issues leveraging heterogeneous attributes compounded density map. We identify three geometric/semantic/numeric essentially important estimation, demonstrate how effectively utilize...

To realize trajectory prediction, most previous methods adopt the parameter-based approach, which encodes all seen past-future instance pairs into model parameters. However, in this way, parameters come from instances, means a huge amount of irrelevant instances might also involve predicting current situation, disturbing performance. provide more explicit link between situation and we imitate mechanism retrospective memory neuropsychology propose MemoNet, an instance-based approach that...

Image quality assessment (IQA) has been an active research area during last decades. Many existing objective IQA models share a similar two-step structure with measuring local distortion before pooling. Compared the rapid development for measurement, seldom effort made dedicated to effective pooling schemes. In this paper, we design new model via analysis of distribution affected by image content and distortion. That is, distributions position, intensity, frequency changes, histogram changes...

Viewing distance and image resolution have substantial influences on quality assessment (IQA), but this issue has been highly overlooked in the literature so far. In paper, we examine problem of optimal adjustment as a preprocessing step for IQA. general, sampling visual information by human eyes' optics is approximately low-pass process. For given scene, amount extractable greatly depends viewing resolution. We first introduce novel dedicated distance-changed database (VDID2014) with two...

Input binarization has shown to be an effective way for network acceleration. However, previous scheme could regarded as simple pixel-wise thresholding operations (i.e., order-one approximation) and suffers a big accuracy loss. In this paper, we propose high-order scheme, which achieves more accurate approximation while still possesses the advantage of binary operation. particular, proposed recursively performs residual quantization yields series input images with decreasing magnitude...