Medical image datasets often exhibit severe long-tailed distributions, making rare disease recognition challenging. Existing models struggle to balance head- and tail-class performance: single-expert models lack flexibility, while multi-expert methods rely on static aggregation without explicit specialization. We propose an uncertainty-aware expert routing framework that integrates explicit expert specialization with adaptive aggregation guided by epistemic uncertainty. Three experts are dedicated to common, balanced, and rare disease groups, respectively, while an Evidential Deep Learning (EDL) branch estimates uncertainty to guide a lightweight router in dynamically weighting expert outputs. This design allows ambiguous cases to receive greater influence from the rare-disease expert. Experiments on ISIC 2019 and HAM10000 demonstrate state-of-the-art results, notably improving tail-class accuracy and calibration while maintaining strong overall performance, validating the effectiveness of uncertainty-driven expert routing for fair and reliable medical image classification.

The size of deep learning models in artificial intelligence (AI) software is increasing rapidly, which hinders the large-scale deployment on resource-restricted devices (e.g., smartphones). To mitigate this issue, AI software compression plays a crucial role, which aims to compress model size while keeping high performance. However, the intrinsic defects in the big model may be inherited by the compressed one. Such defects may be easily leveraged by attackers, since the compressed models are usually deployed in a large number of devices without adequate protection. In this paper, we try to address the safe model compression problem from a safety-performance co-optimization perspective. Specifically, inspired by the test-driven development (TDD) paradigm in software engineering, we propose a test-driven sparse training framework calledSafeCompress. By simulating the attack mechanism as the safety test, SafeCompress can automatically compress a big model to a small one following the dynamic sparse training paradigm. Then, considering two kinds of representative and heterogeneous attack mechanisms i.e., black-box membership inference attack and white-box membership inference attack, we develop two concrete instances called BMIA-SafeCompress and WMIA-SafeCompress. Further, we implement another instance called MMIA-SafeCompress by extending SafeCompress to defend against the occasion when attackers conduct black-box and white-box membership inference attacks simultaneously. Extensive experiments are conducted on five datasets for both computer vision and natural language processing tasks. The results verify the effectiveness and generalizability of our method. We also discuss how to adapt SafeCompress to other attacks besides MIA, demonstrating the flexibility of SafeCompress.

Objective image quality assessment (IQA) aims to evaluate image quality consistently with human perception. Most of the existing perceptual IQA metrics cannot accurately represent the degradations from different types of distortion, e.g., existing structural similarity metrics perform well on content-dependent distortions while not as well as peak signal-to-noise ratio (PSNR) on content-independent distortions. In this paper, we integrate the merits of the existing IQA metrics with the guide of the recently revealed internal generative mechanism (IGM). The IGM indicates that the human visual system actively predicts sensory information and tries to avoid residual uncertainty for image perception and understanding. Inspired by the IGM theory, we adopt an autoregressive prediction algorithm to decompose an input scene into two portions, the predicted portion with the predicted visual content and the disorderly portion with the residual content. Distortions on the predicted portion degrade the primary visual information, and structural similarity procedures are employed to measure its degradation; distortions on the disorderly portion mainly change the uncertain information and the PNSR is employed for it. Finally, according to the noise energy deployment on the two portions, we combine the two evaluation results to acquire the overall quality score. Experimental results on six publicly available databases demonstrate that the proposed metric is comparable with the state-of-the-art quality metrics.

Reduced-reference (RR) image quality assessment (IQA) aims to use less data about the reference image and achieve higher evaluation accuracy. Recent research on brain theory suggests that the human visual system (HVS) actively predicts the primary visual information and tries to avoid the residual uncertainty for image perception and understanding. Therefore, the perceptual quality relies to the information fidelities of the primary visual information and the residual uncertainty. In this paper, we propose a novel RR IQA index based on visual information fidelity. We advocate that distortions on the primary visual information mainly disturb image understanding, and distortions on the residual uncertainty mainly change the comfort of perception. We separately compute the quantities of the primary visual information and the residual uncertainty of an image. Then the fidelities of the two types of information are separately evaluated for quality assessment. Experimental results demonstrate that the proposed index uses few data (30 bits) and achieves high consistency with human perception.

In this paper, we introduce a novel just noticeable difference (JND) estimation model based on the unified brain theory, namely the free-energy principle. The existing pixel-based JND models mainly consider the orderly factors and always underestimate the JND threshold of the disorderly region. Recent research indicates that the human visual system (HVS) actively predicts the orderly information and avoids the residual disorderly uncertainty for image perception and understanding. Thus, we suggest that there exists disorderly concealment effect which results in high JND threshold of the disorderly region. Beginning with the Bayesian inference, we deduce an autoregressive model to imitate the active prediction of the HVS. Then, we estimate the disorderly concealment effect for the novel JND model. Experimental results confirm that the proposed JND model outperforms the relevant existing ones. Furthermore, we apply the proposed JND model in image compression, and around 15% of bit rate can be reduced without jeopardizing the perceptual quality.

In this paper, we propose a new image quality assessment (IQA) scheme, with emphasis on gradient similarity. Gradients convey important visual information and are crucial to scene understanding. Using such information, structural and contrast changes can be effectively captured. Therefore, we use the gradient similarity to measure the change in contrast and structure in images. Apart from the structural/contrast changes, image quality is also affected by luminance changes, which must be also accounted for complete and more robust IQA. Hence, the proposed scheme considers both luminance and contrast-structural changes to effectively assess image quality. Furthermore, the proposed scheme is designed to follow the masking effect and visibility threshold more closely, i.e., the case when both masked and masking signals are small is more effectively tackled by the proposed scheme. Finally, the effects of the changes in luminance and contrast-structure are integrated via an adaptive method to obtain the overall image quality score. Extensive experiments conducted with six publicly available subject-rated databases (comprising of diverse images and distortion types) have confirmed the effectiveness, robustness, and efficiency of the proposed scheme in comparison with the relevant state-of-the-art schemes.

Objective video quality assessment (VQA) is the use of computational models to evaluate the video quality in line with the perception of the human visual system (HVS). It is challenging due to the underlying complexity, and the relatively limited understanding of the HVS and its intricate mechanisms. There are three important issues that arise in objective VQA in comparison with image quality assessment: 1) the temporal factors apart from the spatial ones also need to be considered, 2) the contribution of each factor (spatial and temporal) and their interaction to the overall video quality need to be determined, and 3) the computational complexity of the resultant method. In this paper, we seek to tackle the first issue by utilizing the worst case pooling strategy and the variations of spatial quality along the temporal axis with proper analysis and justification. The second issue is addressed by the use of machine learning; we believe this to be more convincing since the relationship between the factors and the overall quality is derived via training with substantial ground truth (i.e., subjective scores). Experiments conducted using publicly available video databases show the effectiveness of the proposed full-reference (FR) algorithm in comparison to the relevant existing VQA schemes. Focus has also been placed on demonstrating the robustness of the proposed method to new and untrained data. To that end, cross-database tests have been carried out to provide a proper perspective of the performance of proposed scheme as compared to other VQA methods. The third issue regarding the computational costs also plays a key role in determining the feasibility of a VQA scheme for practical deployment given the large amount of data that needs to be processed/analyzed in real time. A limitation of many existing VQA algorithms is their higher computational complexity. In contrast, the proposed scheme is more efficient due to its low complexity without jeopardizing the prediction accuracy.

All existing video coding standards developed so far deem video as a sequence of natural frames (formed in the XY plane), and treat spatial redundancy (redundancy along X and Y directions) and temporal redundancy (redundancy along T direction) differently and separately. In this paper, we investigate into a new compression (redundancy reduction) method for video in which the frames are allowed to be formed in a non-XY plane. We are to exploit fuller extent of video redundancy, and propose an adaptive optimal compression plane determination process to be used as a preprocessing step prior to any standard video coding scheme. The essence of the scheme is to form the frames in the plane formed by two axes (among X, Y, and T) corresponding to signal correlation evaluation, which enables better prediction (therefore better compression). In spite of the simplicity of the proposed method, it can be used for both lossless and lossy compression, and with and without interframe prediction. Extensive experimental results show that the new coding method improves the performance of the video coding for a number of coding methods (inclusive of lossless and near-lossless Motion JPEG-LS, Motion JPEG, Motion JPG2K, H.264 intraonly profile, and H.264) and videos with different visual content.

Contrast masking (CM) on edge and textured regions have to be distinguished since distortions on edge regions are easier to be noticed than that on textured regions. Therefore, how to efficiently estimate the CM on edge and textured regions of an image is a key issue for accurate JND (Just Noticeable Difference) estimation. An enhanced image domain JND estimator is devised in this paper with new model for CM. We use the total variation method to obtain a structural image (which contains edge information) and a textural image (which contains texture information) from the input image, and then evaluate the CM for the two images separately rather than the whole image, and hence edge and texture are better distinguished and the under-estimation of JND on textured regions can be effectively avoided. Experimental results of subjective viewing confirm that the proposed model is capable of determining more accurate visibility thresholds.

This paper presents a framework for video coding, which is compatible with the existing H.264/AVC standard (for preprocessing). Since the video sequences are of 3D data matrices and the traditional XY image frame plane is not always the best for video coding in terms of rate-distortion (RD) performance, optimal compression plane (OCP) determination models are developed to improve the coding efficiency. We design coding plane level RD cost function (in analog to the one in H.264 macroblock level) to measure the RD performance of each coding plane. We jointly consider the available computational resource and the RD performance in order to determine an adaptive OCP. The performance of the proposed coding framework is assessed with a number video sequences. Extensive experimental results show that the new coding framework can achieve better RD performance at approximately the same computational complexity, for videos with different visual content.

In just noticeable difference (JND) models, evaluation of contrast masking (CM) is a crucial step. More specifically, CM due to edge masking (EM) and texture masking (TM) needs to be distinguished due to the entropy masking property of the human visual system. However, TM is not estimated accurately in the existing JND models since they fail to distinguish TM from EM. In this letter, we propose an enhanced pixel domain JND model with a new algorithm for CM estimation. In our model, total-variation based image decomposition is used to decompose an image into structural image (i.e., cartoon like, piecewise smooth regions with sharp edges) and textural image for estimation of EM and TM, respectively. Compared with the existing models, the proposed one shows its advantages brought by the better EM and TM estimation. It has been also applied to noise shaping and visual distortion gauge, and favorable results are demonstrated by experiments on different images.

In this paper, we investigate into a new lossless compression framework for video. We are to exploit fuller scale of temporal redundancy, and propose an adaptive optimal compression plane (OCP) determination process to be used with the standard intra-frame lossless predictive coding scheme. The essence of the scheme is to perform the prediction on the plane formed by two axes (among X, Y and T) corresponding to signal correlation evaluation, which enables better prediction (therefore better compression). In spite of the simplicity of the proposed method, simulation results show that the rate saving of the resultant system is 27.0% on average with respect to JPEG-LS, for a large number of videos with different visual content.