Image deraining is crucial for vision applications but is challenged by the complex multi-scale physics of rain and its coupling with scenes. To address this challenge, a novel approach inspired by multi-stage image restoration is proposed, incorporating Point Spread Function (PSF) mechanisms to reveal the image degradation process while combining dynamic physical modeling with sequential feature fusion transfer, named SD-PSFNet. Specifically, SD-PSFNet employs a sequential restoration architecture with three cascaded stages, allowing multiple dynamic evaluations and refinements of the degradation process estimation. The network utilizes components with learned PSF mechanisms to dynamically simulate rain streak optics, enabling effective rain-background separation while progressively enhancing outputs through novel PSF components at each stage. Additionally, SD-PSFNet incorporates adaptive gated fusion for optimal cross-stage feature integration, enabling sequential refinement from coarse rain removal to fine detail restoration. Our model achieves state-of-the-art PSNR/SSIM metrics on Rain100H (33.12dB/0.9371), RealRain-1k-L (42.28dB/0.9872), and RealRain-1k-H (41.08dB/0.9838). In summary, SD-PSFNet demonstrates excellent capability in complex scenes and dense rainfall conditions, providing a new physics-aware approach to image deraining.

Large language models (LLMs) still struggle with multi-hop reasoning over knowledgegraphs (KGs), and we identify a previously overlooked structural reason for this difficulty: Transformer attention heads naturally specialize in distinct semantic relations across reasoning stages, forming a hop-aligned relay pattern.This key finding suggests that multi-hop reasoning is inherently multi-view, yet existing KG-based retrieval-augmented generation (KG-RAG) systems collapse all reasoning hops into a single representation, flat embedding space, suppressing this implicit structure and causing noisy or drifted path exploration.We introduce ParallaxRAG, a symmetric multi-view framework that decouples queries and KGs into aligned, head-specific semantic spaces.By enforcing relational diversity across multiple heads while constraining weakly related paths, ParallaxRAG constructs more accurate, cleaner subgraphs and guides LLMs through grounded, hop-wise reasoning.On WebQSP and CWQ, it achieves state-of-the-art retrieval and QA performance, substantially reduces hallucination, and generalizes strongly to the biomedical BioASQ benchmark.Our implementation is available at https://github.com/ LucaLiu1313/ParallaxRAG.

The rapid growth of medical imaging data presents significant challenges in diagnostic accuracy, data privacy, and computational efficiency. Traditional centralized AI models struggle with scalability and pose risks to patient confidentiality due to data aggregation. Moreover, heterogeneous medical data across institutions complicates the development of robust diagnostic tools. To address these issues, we propose MID-LLM, a novel framework that integrates Large Language Models (LLMs) with a blockchain-based federated learning system for medical image analysis. It also ensures the security and privacy of sensitive medical data across decentralized networks. MID-LLM uses verification mechanisms to ensure the global model’s integrity. It also employs aggregation techniques to reduce bias and improve training efficiency. Experiments on the BraTS 2020 dataset show that MID-LLM outperforms traditional federated learning, achieving higher Dice scores with improved computational efficiency. These results highlight MID-LLM’s potential to enhance diagnostic accuracy while offering a scalable, secure solution for AI in healthcare.

Convolutional neural networks (CNNs) have been successfully applied to various fields. However, CNNs' overparameterization requires more memory and training time, making it unsuitable for some resource-constrained devices. To address this issue, filter pruning as one of the most efficient ways was proposed. In this article, we propose a feature-discrimination-based filter importance criterion, uniform response criterion (URC), as a key component of filter pruning. It converts the maximum activation responses into probabilities and then measures the importance of the filter through the distribution of these probabilities over classes. However, applying URC directly to global threshold pruning may cause some problems. The first problem is that some layers will be completely pruned under global pruning settings. The second problem is that global threshold pruning neglects that filters in different layers have different importance. To address these issues, we propose hierarchical threshold pruning (HTP) with URC. It performs a pruning step limited in a relatively redundant layer rather than comparing the filters' importance across all layers, which can avoid some important filters being pruned. The effectiveness of our method benefits from three techniques: 1) measuring filter importance by URC; 2) normalizing filter scores; and 3) conducting prune in relatively redundant layers. Extensive experiments on CIFAR-10/100 and ImageNet show that our method achieves the state-of-the-art performance on multiple benchmarks.

Graph convolutional networks (GCNs) are a popular approach to learn the feature embedding of graph-structured data, which has shown to be highly effective as well as efficient in performing node classification in an inductive way. However, with massive nongraph-organized data existing in application scenarios nowadays, it is critical to exploit the relationships behind the given groups of data, which makes better use of GCN and broadens the application field. In this article, we propose the f uzzy g raph s ubspace c onvolutional n etwork (FGSCN) to provide a brand-new paradigm for feature embedding and node classification with graph convolution (GC) when given an arbitrary collection of data. The FGSCN performs GC on the f uzzy s ubspace ($\mathcal {F}$-space), which simultaneously learns from the underlying subspace information in the low-dimensional space as well as its neighborliness information in the high-dimensional space. In particular, we construct the fuzzy homogenous graph$\mathcal {G}_{\mathcal {F}}$on the$\mathcal {F}$-space by fusing the homogenous graph of neighborliness$\mathcal {G}_{\mathcal {N}}$and homogenous graph of subspace$\mathcal {G}_{\mathcal {S}}$(defined by the affinity matrix of the low-rank representation). Here, it is proven that the GC on$\mathcal {F}$-space will propagate both the local and global information through fuzzy set theory. We evaluated FGSCN on 15 unique datasets with different tasks (e.g., feature embedding, visual recognition, etc.). The experimental results showed that the proposed FGSCN has significant superiority compared with current state-of-the-art methods.

Abstract Point cloud completion is prevalent due to the insufficient results from current point cloud acquisition equipments, where a large number of point data failed to represent a relatively complete shape. Existing point cloud completion algorithms, mostly encoder‐decoder structures with grids transform (also presented as folding operation), can hardly obtain a persuasive representation of input clouds due to the issue that their bottleneck‐shape result cannot tell a precise relationship between the global and local structures. For this reason, this article proposes a novel point cloud completion model based on a Stack‐Style Folding Network (SSFN). Firstly, to enhance the deep latent feature extraction, SSFN enhances the exploitation of shape feature extractor by integrating both low‐level point feature and high‐level graphical feature. Next, a precise presentation is obtained from a high dimensional semantic space to improve the reconstruction ability. Finally, a refining module is designed to make a more evenly distributed result. Experimental results shows that our SSFN produces the most promising results of multiple representative metrics with a smaller scale parameters than current models.

The sparsity is an attractive property that has been widely and intensively utilized in various image processing fields (e.g., robust image representation, image compression, image analysis, etc.). Its actual success owes to the exhaustive mining of the intrinsic (or homogenous) information from the whole data carrying redundant information. From the perspective of image representation, the sparsity can successfully find an underlying homogenous subspace from a collection of training data to represent a given test sample. The famous sparse representation (SR) and its variants embed the sparsity by representing the test sample using a linear combination of training samples with $L_{0}$ -norm regularization and $L_{1}$ -norm regularization. However, although these state-of-the-art methods achieve powerful and robust performances, the sparsity is not fully exploited on the image representation in the following three aspects: 1) the within-sample sparsity, 2) the between-sample sparsity, and 3) the image structural sparsity. In this paper, to make the above-mentioned multi-context sparsity properties agree and simultaneously learned in one model, we propose the concept of consensus sparsity (Con-sparsity) and correspondingly build a multi-context sparse image representation (MCSIR) framework to realize this. We theoretically prove that the consensus sparsity can be achieved by the $L_{\infty }$ -induced matrix variate based on the Bayesian inference. Extensive experiments and comparisons with the state-of-the-art methods (including deep learning) are performed to demonstrate the promising performance and property of the proposed consensus sparsity.

The Collaborative Representation (CR) framework has provided various effective and efficient solutions to pattern analysis. By leveraging between discriminative coefficient coding (l 2 regularization) and the best reconstruction quality (collaboration), the CR framework can exploit discriminative patterns efficiently in high-dimensional space. Due to the limitations of its linear representation mechanism, the CR must sacrifice its superior efficiency for capturing the non-linear information with the kernel trick. Besides this, even if the coding is indispensable, there is no mechanism designed to keep the CR free from inevitable noise brought by real-world information systems. In addition, the CR only emphasizes exploiting discriminative patterns on coefficients rather than on the reconstruction. To tackle the problems of primitive CR with a unified framework, in this article we propose the Euler Collaborative Representation (E-CR) framework. Inferred from the Euler formula, in the proposed method, we map the samples to a complex space to capture discriminative and non-linear information without the high-dimensional hidden kernel space. Based on the proposed E-CR framework, we form two specific classifiers: the Euler Collaborative Representation based Classifier (E-CRC) and the Euler Probabilistic Collaborative Representation based Classifier (E-PROCRC). Furthermore, we specifically designed a robust algorithm for E-CR (termed as R-E-CR ) to deal with the inevitable noises in real-world systems. Robust iterative algorithms have been specially designed for solving E-CRC and E-PROCRC. We correspondingly present a series of theoretical proofs to ensure the completeness of the theory for the proposed robust algorithms. We evaluated E-CR and R-E-CR with various experiments to show its competitive performance and efficiency.

Zero-shot learning (ZSL) aims to recognize image instances of unseen classes solely based on the semantic descriptions of the unseen classes. In this field, Generalized Zero-Shot Learning (GZSL) is a challenging problem in which the images of both seen and unseen classes are mixed in the testing phase of learning. Existing methods formulate GZSL as a semantic-visual correspondence problem and apply generative models such as Generative Adversarial Networks and Variational Autoencoders to solve the problem. However, these methods suffer from the bias problem since the images of unseen classes are often misclassified into seen classes. In this work, a novel model named the Dual Projective model for Zero-Shot Learning (DPZSL) is proposed using text descriptions. In order to alleviate the bias problem, we leverage two autoencoders to project the visual and semantic features into a latent space and evaluate the embeddings by a visual-semantic correspondence loss function. An additional novel classifier is also introduced to ensure the discriminability of the embedded features. Our method focuses on a more challenging inductive ZSL setting in which only the labeled data from seen classes are used in the training phase. The experimental results, obtained from two popular datasets—Caltech-UCSD Birds-200-2011 (CUB) and North America Birds (NAB)—show that the proposed DPZSL model significantly outperforms both the inductive ZSL and GZSL settings. Particularly in the GZSL setting, our model yields an improvement up to 15.2% in comparison with state-of-the-art CANZSL on datasets CUB and NAB with two splittings.

Dictionary-based Classification (DC) has been a promising learning theory in multimedia computing. Previous studies focused on learning a discriminative dictionary as well as the sparsest representation based on the dictionary, to cope with the complex conditions in real-world applications. However, robustness by learning only one single dictionary is far from the optimal level. What is worse, it cannot take advantage of the available techniques proven in modern machine learning, like data augmentation, to mitigate the same problem. In this work, we propose a novel method that utilizes joint Augmented and Compressed Dictionaries for Robust Dictionary-based Classification (ACD-RDC). For optimization under the noise model introduced by real-world conditions, the objective function of ACD-RDC incorporates only two simple, but well-designed constraints, including one enhanced sparsity constraint by the general data augmentation, which requires less case-by-case and sophisticated tuning, and another discriminative constraint solved by a jointly learned dictionary. The optimization of the objective function is then deduced theoretically to an approximate linear problem. The sparsity and discrimination enhanced by data augmentation guarantees the robustness for image classification under various conditions, which constructs the first positive case using data augmentation to obtain robust dictionary-based classification. Numerous experiments have been conducted on popular facial and object image datasets. The results demonstrate that ACD-RDC obtains more promising classification on diversely collected images than the current dictionary-based classification methods. ACD-RDC is also confirmed to be a state-of-the-art classification method when using deep features as inputs.

Deep learning methods have widely been applied to accurately identify and segment individual teeth in panoramic X-ray radiographs. However, the task becomes challenging as deep learning models grow deeper and wider. Contextual information have to pass through many layers leading to features vanishing before reaching the end of the model. This study proposes a deep learning-based three-step model to identify and segment individual teeth from panoramic X-ray radiographs to address the issues above. Firstly, an automatic binarized transformation of panoramic images to deal with computational complexity of high-dimensionality with respect to less training data is conducted. The transformed images are then trained on with a vanilla top-down learnable feature connection based UNet. Specifically, a learnable feature connection module that incorporates an improved squeeze and excitation module with dense connections to ensure feature propagation and facilitating information flow throughout the network is designed and conveniently plunged into a UNet. Finally, accurate individual tooth identification and segmentation is achieved using proposed regions via bounding detection techniques. Extensive evaluation on a publicly available dental panoramic X-ray benchmark demonstrated the effectiveness the proposed scheme by obtaining a Dice score of 96.94%, 97.93% for accuracy, 97.03% for recall, 96.81% for precision, 97.61% for specificity and 93.52% for Jaccard Coefficient Score, which are significantly superior to the recent state-of-the-art methods for individual tooth identification and segmentation.

Latent subspace learning aims to produce a latent representation for better reconstruction and classification from high-dimensional data through exploiting the optimal subspace. Current latent subspace learning methods commonly have three problems: 1) The discriminative property is ignored when learning the latent subspace, 2) The redundancy exists between the latent subspace and the prediction space, 3) There is no unified latent subspace that exploits knowledge jointly from the raw space, latent subspace, and label space. In this paper, we formulate theJointDiscriminativeLatentSubspaceLearning (JDLSL) problem to address these issues, and provide its optimization solution. JDLSL learns image representation from two aspects: a) the joint learning of latent subspaces for data reconstruction and prediction, b) the joint learning of label space and latent subspace for data reconstruction. To integrate knowledge from the joint learning, we organize the sparsity-induced latent subspace, where row-sparsity and column sparsity are simultaneously imposed. We provide the theoretical proof for the discriminativity learning ability of the sparsity-induced latent subspace. Extensive experiments and comparisons with the state-of-the-art showed that the proposed method has better performance. JDLSL shows a competitive performance with deep features compared to deep learning architectures, reflecting it potential integrating with deep learning.

Image classification is a fundamental component in modern computer vision systems, where sparse representation-based classification has drawn a lot of attention due to its robustness. However, on the optimization of sparse learning systems, regularization and data augmentation are both powerful, but currently isolated. We believe that regularization and data augmentation can cooperate to generate a breakthrough in robust image classification. In this article, we propose a novel framework, regularization on augmented data (READ), which creates diversification in the data using the generic augmentation techniques to implement robust sparse representation-based image classification. When the training data are augmented, READ applies a distinct regularizer,$l_{1}$or$l_{2}$, in particular, on the augmented training data apart from the original data, so that regularization and data augmentation are utilized and enhanced synchronously. We introduce an elaborate theoretical analysis on how to optimize the sparse representation by both$l_{1}$-norm and$l_{2}$-norm with the generic data augmentation and demonstrate its performance in extensive experiments. The results obtained on several facial and object datasets show that READ outperforms many state-of-the-art methods when using deep features.

In a typical Generative Adversarial Network (GAN), a noise is sampled to generate fake samples via a series of convolutional operations after random initialization. However, current GANs merely rely on the pixel space to sample the noise, in-creasing the difficulty of approaching the target distribution. Fortunately, the long proven wavelet transformation is able to decompose multiple spectral information from the images. In this work, we propose a novel multi-channel wavelet-based filtering method for GANs, to cope with this problem. The proposed WaveletNet embeds a wavelet deconvolution layer in the generator to take advantage of the wavelet deconvolution. By learning a filter with multiple channels, or multiple convolutional filters, it can efficiently homogenize the sampled noise via an averaging operation, to generate high-fidelity samples. We conducted benchmark experiments on the Fashion-MNIST, KMNIST, and SVHN datasets through an open GAN bench-mark tool. The results showed that WaveletGAN has excellent performance in generating high-fidelity samples.

Representation-based classification (RBC) has been attracting a great deal of attention in pattern recognition. As a typical extension to RBC, collaborative representation-based classification (CRC) has demonstrated its superior performance in various image classification tasks. Ideally, we expect that the learned class-specific representations for a testing sample are discriminative, and the representation computed for the true class dominates the final representation of the testing sample. Most existing CRC-based methods can learn pattern discrimination, but cannot differentiate the contribution of class-specific representations to the classification of each testing sample. It is challenging for a representation-based classifier to retain both properties. To address this challenge and further improve CRC's classification performance, we propose a novel CRC-based method, class mean-weighted discriminative collaborative representation-based classifier (CMW-DCRC). Its objective function penalises the standard l 2 -norm residuals with two discriminative regularisation terms. A decorrelating term makes the class-specific representations more discriminative, and a newly designed class mean-weighted term that promotes the training samples from individual classes to competitively reconstruct the testing sample while boosting the contribution of the true class. To further enhance the robustness of CRC, we extend CMW-DCRC by replacing the l2-norm coding residual with a l1-norm coding residual, and solve the optimisation problem with an iteratively reweighted least square algorithm. Extensive experimental results on nine image data sets have shown that our methods outperform the state-of-the-art RBC-based methods.

Dictionary-based classification has been promising in knowledge discovery from image data, due to its good performance and interpretable theoretical system. Dictionary learning effectively supports both small- and large-scale datasets, while its robustness and performance depends on the atoms of the dictionary most of the time. Empirically, using a large number of atoms is helpful to obtain a robust classification, while robustness cannot be ensured when setting a small number of atoms. However, learning a huge dictionary dramatically slows down the speed of classification, which is especially worse on the large-scale datasets. To address the problem, we propose a Fast and Robust Dictionary-based Classification (FRDC) framework, which fully utilizes the learned dictionary for classification by staging - and -norms to obtain a robust sparse representation. The new objective function, on the one hand, introduces an additional -norm term upon the conventional -norm optimization, which generates a more robust classification. On the other hand, the optimization based on both - and -norms is solved in two stages, which is much easier and faster than current solutions. In this way, even when using a limited size of dictionary, which makes sure the classification runs very fast, it still can gain higher robustness for multiple types of image data. The optimization is then theoretically analyzed in a new formulation, close but distinct to elastic-net, to prove it is crucial to improve the performance under the premise of robustness. According to our extensive experiments conducted on four image datasets for face and object classification, FRDC keeps generating a robust classification no matter whether using a small or large number of atoms. This guarantees a fast and robust dictionary-based image classification. Furthermore, when simply using deep features extracted via some popular pre-trained neural networks, it outperforms many state-of-the-art methods on the specific datasets.

Deep convolutional neural networks provide a powerful feature learning capability for image classification. The deep image features can be utilized to deal with many image understanding tasks like image classification and object recognition. However, the robustness obtained in one dataset can be hardly reproduced in the other domain, which leads to inefficient models far from state-of-the-art. We propose a deep collaborative weight-based classification (DeepCWC) method to resolve this problem, by providing a novel option to fully take advantage of deep features in classic machine learning. It firstly performs the $l_2$-norm based collaborative representation on the original images, as well as the deep features extracted by deep CNN models. Then, two distance vectors, obtained based on the pair of linear representations, are fused together via a novel collaborative weight. This collaborative weight enables deep and classic representations to weigh each other. We observed the complementarity between two representations in a series of experiments on 10 facial and object datasets. The proposed DeepCWC produces very promising classification results, and outperforms many other benchmark methods, especially the ones claimed for Fashion-MNIST. The code is going to be published in our public repository\footnote{https://github.com/zengsn/research}.

Sparse representation is a significant method to perform image classification for face recognition. Sparsity of the image representation is the key factor for robust image classification. As an improvement to sparse representation-based classification, collaborative representation is a newer method for robust image classification. Training samples of all classes collaboratively contribute together to represent one single test sample. The ways of representing a test sample in sparse representation and collaborative representation are very different, so we propose a novel method to integrate both sparse and collaborative representations to provide improved results for robust face recognition. The method first computes a weighted average of the representation coefficients obtained from two conventional algorithms, and then uses it for classification. Experiments on several benchmark face databases show that our algorithm outperforms both sparse and collaborative representation-based classification algorithms, providing at least a 10% improvement in recognition accuracy.