Abstract The exponential growth of multimedia data in today’s highly digitized, mobile-centric society poses formidable challenges to efficient data storage and rapid information retrieval. To overcome these bottlenecks, we introduce ViT-LSH, a highly optimized algorithm that seamlessly integrates Vision Transformer (ViT) with locality-sensitive hashing (LSH) to deliver unprecedented image matching accuracy and retrieval speeds. The superiority of ViT-LSH stems from its dual-engine architecture. First, the ViT module revolutionizes feature extraction by dividing images into fixed-size patches and projecting them into token sequences without relying on down-sampling. This perfectly preserves the original image resolution and achieves exceptional global information modeling for semantic segmentation. Second, the LSH module drastically minimizes space consumption and accelerates query processing by mapping proximate data points to identical hash values. Evaluated against five existing baseline schemes, our LSH approach demonstrates vastly superior time and space efficiency while guaranteeing highly accurate approximate nearest neighbor queries. The proposed algorithm follows a streamlined, robust pipeline: image preprocessing (e.g., denoising) and segmentation, followed by ViT-driven feature extraction, dimensionality reduction, and LSH-based hash coding. By utilizing Hamming distance for direct similarity calculation, the system achieves ultra-fast sorting and matching. Extensive experimental results unequivocally validate that ViT-LSH provides a massive leap forward in both the computational efficiency and precision of large-scale image matching.

The YOLO series has achieved remarkable progress in real-time object detection, yet nano-scale variants remain constrained by single-round feature pyramid topologies where the global correlation prior is distributed only once and extreme pyramid levels exchange information indirectly. In this paper, we propose YOLO-Prism, an efficient nano-scale detector built upon YOLOv13 that addresses this single-round limitation through asymmetric multi-round refinement. We propose a Multi-Round Feature Pyramid Network (MR-FPN) that exploits the decoupling between YOLOv13’s precomputed correlation prior and its lightweight gated distribution mechanism, appending a secondary top-down refinement pass at only ∼6% parameter overhead under a “compute once, distribute many” paradigm. Subsequently, as a near-zero-cost complement, we selectively integrate DCT-based Multi-Spectral Channel Attention (MSCA) at gradient-identified critical layers, contributing an additional gain at merely 0.01M parameters. Extensive experiments on MS COCO, PAS-CAL VOC, and VisDrone benchmarks demonstrate that YOLO-xsPrism achieves 41.66% AP50:95 on COCO val2017, surpassing the strong YOLOv13n baseline by 0.62 points under identical training conditions and outperforming YOLOv12n by 1.10 points, at 2.63M parameters and 7.1 GFLOPs. On YOLOv11n, whose neck lacks such a prior, the same asymmetric topology yields 0.50 AP50:95 at 3.8% overhead, confirming that the benefit of multi-round refinement is independent of any specific gating mechanism. This lightweight algorithm can be applied to intelligent video measurement, industrial edge detection sensors, and IoT monitoring.

Drawing inspiration from the brain's neurocognitive mechanisms of information chunking and topographic mapping, adaptive decision-making requires neural-grounded architectures that are interpretable and resilient to uncertainty. In this paper, we propose a novel boosted fuzzy manifold granule hypersurface classifier (BFMGHC). The algorithm performs classification at the “information granule" level, realizing an intelligent modeling method that is closer to human cognition, more interpretable, and more accommodating of uncertainty. The classifier mainly consists of three main parts: (1) A manifold-based measurement method for samples that preserves local topological structure is designed, echoing the topographic representations in neural dynamics. Based on this, a global optimization clustering algorithm is proposed and integrated with the Dask framework to achieve scalable hierarchical parallel granulation from raw inputs to high-level semantic granules. (2) In the fuzzy manifold granule space, a measurement method and a hypersurface classifier are constructed, utilizing a particle swarm optimization method for parameter solving. (3) To improve interpretability, weights are assigned to different granules and base classifiers, resembling bio-inspired neuromodulation to ensure stable behavior. The proposed BFMGHC was verified on three financial risk assessment datasets in the UCI Machine Learning Repository (Default of Credit Card Clients, Bank Marketing, and German Credit Data) and achieved superior performance.

Action Quality Assessment (AQA)—the task of evaluating how well human actions are performed—is essential in domains such as sports and medicine. Existing AQA methods typically rely on score regression following feature extraction but often neglect the ambiguity inherent in extracted features. In this work, we introduce a novel AQA framework that incorporates a modified attention module to better capture relevant information. Our approach segments video data into clips, extracts features using the I3D network, and applies attention mechanisms to highlight salient features while suppressing irrelevant ones. To assess feature quality, we employ score distribution regression and propose an uncertainty-aware score distribution learning strategy that models features as Gaussian distributions. We further leverage Variational Autoencoders (VAEs) to capture complex latent representations and quantify uncertainty. Extensive experiments on the MTL-AQA and JIGSAWS datasets demonstrate the effectiveness and robustness of our proposed method.

With the continuous improvement of hardware device performance and the development of artificial intelligence technology, radar point clouds have become an important component in related fields such as autonomous driving due to their high anti-interference ability. However, existing detection technologies still pose a challenge in identifying small objects. The algorithm first adopts a feature pyramid structure to address the size difference between detected targets in actual 3D object detection scenes. Through multi-scale fusion, the model can retain more delicate features during the encoding stage. In the feature extraction stage, a cascade approach of deformable convolution and self-attention mechanism was adopted. The deformable convolution can adaptively change the shape and size of the input features to filter out noise effects. The self-attention mechanism is used to consider the global information of all pixels in the image, helping the model better understand the image content and improve the accuracy of object detection. The experimental results show that the proposed algorithm performs well in simple, medium, and difficult levels in BEV mode, 3D mode, and AOS mode. Compared with the original algorithm, the average accuracy has been significantly improved, especially in the difficulty level of BEV mode, where the improvement rates for cars, pedestrians, and bicycles reach 5.75%, 3.56%, and 7.05%, respectively.

In this work, we design a boosted stochastic fuzzy granular hypersurface classifier (BSFGHC) to resolve the classification issue of numerical data and non-numerical data (such as information granules) from the standpoint of granular computing. The scheme is divided into three parts: first, we present an adaptive cluster center clustering (ACCC) algorithm to achieve cluster centers of the data and to realize the fuzzy granulation of data parallelly based on Spark, which dramatically improves the granulation efficiency; second, we build a fuzzy granular space, design various fuzzy granular operators and measurement in the space to construct fuzzy granular hypersurfaces, create the loss function, and employ Particle Swarm Optimization (PSO) to resolve the optimal fuzzy granular hypersurfaces; third, we randomly divide the fuzzy granules to train multiple optimal fuzzy granular hypersurfaces and combine with the classification accuracy of fuzzy hypersurfaces and the difficulty of fuzzy granule subset classification to form a boosted fuzzy hypersurface to predict the data comprehensively. Experimental results and theoretical analysis demonstrate the outstanding performance of the method.

The traditional 2D convolution approaches are inappropriate for capturing information about local objects and their scale changes. To better depict 3D structures, the fundamental concept behind it is to mix depth information with RGB data to understand 3D space via depth estimations for creating depth maps of 2D images. However, it places a great deal of reliance on the precision of depth estimate, which results in subpar performance. In this paper, a monocular 3D object detection algorithm is proposed to resolve the issues mentioned earlier. The algorithm depends on the estimation of depth, which is essential for accurately detecting objects using a single camera. To begin with, a Normalization-based Attention Module has been deployed to a backbone network with two branches to extract RGB and depth maps separately. To merge the RGB and depth map features, a depth-guided dynamic local convolutional module is subsequently implemented. Depth map learns exemplar kernels from RGB images using a weight learning submodule and dilation convolution. Then it acts locally on the corresponding RGB images to solve the scale-sensitive problem of objects. By utilizing this approach, it becomes possible to extract feature representations from 2D images and map them onto 3D space, minimizing the disparity between the 2D and 3D representations. Experiments indicate that under the moderate setting of the class car - which is the most significant level of the KITTI dataset - our method outperforms other cutting-edge algorithms.

The recognition of traffic signs is a significant challenge for the development of intelligent transportation systems. The small target characteristics of traffic signs during automobile driving make them difficult to recognize, leading to easily missed detection and other issues. To address small target traffic signs, we have made corresponding improvements based on the YOLOv5s algorithm model. Firstly, we enhance the Mosaic online data augmentation in the YOLOv5s algorithm and remove incomplete samples during the training process by setting up a threshold. Secondly, we introduce a small target detection head containing more shallow features, removing the detection head used for detecting large targets. Within the feature pyramid, multi-scale attention feature fusion replaces the original concat features fusion, reducing conflicts in features fusion at different levels. Finally, considering the sensitivity to location deviation when calculating the loss for small targets, we combine the Normalized Wasserstein Distance loss and the Focal EIOU loss as location loss. Experiment findings on the CCTSDB-2021 dataset reveal that our improvements have the most obvious enhancement effect on small target detection, increasing mAP@0.5 (mean Average Precision) by 3.5%.

Focusing on the discovery of the value of in-hospital electronic medical record data for the three "chronic diseases" of diabetes, liver disease and hypertension, it provides data support for improving the hospital's "patient-centered" service level. Through web crawler technology, word frequency analysis technology, WeChat applet development technology, etc., we complete the design and development of big data systems such as data collection, preprocessing, analysis, and visualization, and tap the potential value of ten-year electronic medical record data. The standardized data collation platform and the development of the "Community Online" WeChat applet were completed. The original html data was standardized and stored in a relational database; through data mining, the distribution rules of occupation, age, gender, etc. of regional chronic diseases were found; through word frequency analysis, three kinds of chronic disease admission symptoms, treatment medication and discharge life suggestions were found hot word. Taking the system as the carrier, and through the research on the value discovery of Electronic Medical Records (EMR) data, a systematic chronic disease service system from health warning to admission treatment to discharge tracking has been built for patients with diabetes, liver disease and hypertension. Provide decision-making support for "early warning, early detection, early diagnosis, and early treatment" of chronic diseases and regional improvement of comprehensive management of chronic diseases and scientific treatment.

Abstract The progress of Internet technology and the portability of mobile communication equipment have provided essential convenience for building an information society. Today, with networked and digital platforms as an important way of socializing, the information carrier is more based on images, videos, etc. The explosive growth of these multimedia data poses a huge challenge to efficient data storage and fast information retrieval. We design an algorithm combining Locality-sensitive Hashing (LSH) and Vision Transformer (ViT) to greatly improve the image matching rate and accuracy, and the algorithm is called ViT-LSH. The principle of LSH is to form the same hash value for points with similar distances in the data set. The results show that the LSH algorithm reduces space consumption and improves query processing efficiency compared with the other five schemes. Therefore, the LSH algorithm can help to quickly get the approximate nearest neighbor query results in a probabilistic manner, thereby ensuring the accuracy of the query results and the efficiency of time and space. The ViT divides the image into fixed-size image blocks ( Patch), uses Linear projection to project the Patch flatten to the specified dimension, gets the Tokens sequence, and adopts the sequence as the input of the feature to realize a new segmentation mode. The ViT does not employ down-sampling so that the image resolution is not reduced. The modeling of global information is a new semantic segmentation mode. The steps of ViT-LSH are as follows: first, preprocess the image, such as denoising, then segment the image, then extract features by ViT, and then reduce features' dimension. The LSH performs hash coding on the image features. After calculating the image's hash code, the image's similarity can be directly calculated by the Hamming distance, sorted according to the similarity, and the matching results are obtained. The experimental results exhibit that ViT-LSH can greatly enhance the efficiency and accuracy of image matching.

Convolutional operations extracting effective features have been widely used in the field of deep learning. For the convolution is difficult to process set data, we propose two convolutional operators on fuzzy sets, and build a fuzzy granular classifier. Firstly, a fuzzy granulation is performed on single-atom features of classification systems to form fuzzy conditional granules and fuzzy decision granules. Then, a fuzzy conditional granular vector is constructed from the fuzzy conditional granules, and a convolutional operation is carried out on the granular vector. After that, a new fuzzy feature granule is obtained. The fuzzy feature granule is compared with its corresponding fuzzy decision granule. The result of comparison is back-propagated to the fuzzy granular vector. Simultaneously, weights of the fuzzy granular vector are modified. Thus, a fuzzy granular convolutional classifier is formed by iterating and optimizing the weights of fuzzy granular vectors several times. Furthermore, we prove the difference and derivative of fuzzy granular convolution, which provide a theoretical basis for the back-propagation of the fuzzy granular convolutional classifier. Finally, the convergence effects of the fuzzy granular convolutional operations and the classification performance of the proposed classifier are tested on some UCI datasets. The theoretical analysis and experimental results show that the convolutional operations of fuzzy granular vectors have the characteristics of fast convergence, and the fuzzy granular convolutional classifier obtains a better classification performance.

Using deep learning to colorize grayscale images often requires the preparation of a large number of training images, which also requires very high computing power. In view of this situation, this paper designs a fast colorization method for grayscale images that combines traditional feature extraction with a simple neural network. The method is mainly divided into three steps: Firstly, select the reference color image and train the network, then the target gray image is input into the network to generate the first stage color image; Next, the color image obtained in the previous step is transformed into HSV space, only the V component is retained, and the gray value of the target gray image is used for color synthesis to obtain the second stage color image with clear texture; Finally, the Reinhard algorithm is used for color migration, and then the reference color image is used to color the target image accurately to obtain the third stage color image. Experiments demonstrated that the algorithm proposed is fast, efficient, and robust.

In this paper, we propose a similar image retrieval algorithm based on feature fusion and locality-sensitive hash to address the problems of inadequate representation of image content by individual features and long retrieval time for massive image data. The fusion of global features and attention features makes the image features have both color structure and semantic information, which can better characterize the image content. In the image retrieval stage, the locality-sensitive hash is used to hash encode the image features, the cosine similarity is used as the similarity measure, and finally, the index is built to improve the retrieval efficiency. The similar image retrieval algorithm proposed in this paper has improved the average finding accuracy and recall rate on Caltech 256 and Corel5k datasets compared with other methods, and the retrieval time is greatly reduced.

In recent years, facial image search has been widely used. Although offering considerable convenience, facial image search also poses a severe threat to people’s privacy. How to conduct facial image search while protecting privacy has become a challenge. We design a scheme of searchable encryption facial image. The scheme is separated into two phases: uploading phase and retrieval phase. At the phase of uploading, the facial image is divided into two parts: the obfuscated features of the facial image and the encrypted facial image. In the retrieval phase, the facial image features are extracted and the similarity is calculated with the obfuscated facial features in the cloud after being obfuscated to obtain the closest obfuscated facial features and the corresponding encrypted facial image, then the encrypted facial image is decrypted in order to acquire the original facial image. Specifically, Facenet512 is used to collect facial image features while the obfuscation function is designed to obfuscate those facial image features and the Advanced Encryption Standard (AES) algorithm is used to encrypt facial images. Theoretical analysis and experimental results indicate that the scheme possesses favorable performance and high security.

With the continuous progress of image retrieval technology, in the field of image retrieval, the speed of a search for a desired image from a great deal of image data becomes a hot issue. Convolutional Neural Networks (CNN) have been used in the field of image retrieval. However, many image retrieval systems based on CNN have a poor ability to express image features, resulting in a series of problems such as low retrieval accuracy and robustness. When the target image is retrieved from a large amount of image data, the vector dimension after image coding is high and the retrieval efficiency is low. Locality-sensitive hash is a method to find similar data from massive high latitude data. It reduces the data dimension of the original spatial data through hash coding and conversion, and can also maintain the similarity between the data. The retrieval time and space complexity are low. Therefore, this paper proposes a locality-sensitive hash image retrieval method based on CNN and the attention mechanism. The steps of the method are as follows: using the ResNet50 network as the feature extractor of the image, adding the attention module after the convolution layer of the model, and using the output of the network full connection layer to retrieve the features of the image database, then using the local-sensitive hash algorithm to hash code the image features of the database to reduce the dimension and establish the index, and finally measuring the features of the image to be retrieved and the image database to get the most similar image, completing the content-based image retrieval task. The method in this paper is compared with other image retrieval methods on corel1k and corel5k datasets. The experimental results show that this method can effectively improve the accuracy of image retrieval, and the retrieval efficiency is significantly improved. It also has higher robustness in different scenarios.

Based on the generative adversarial network, this paper proposes a method of colorizing gray images without supervision and applies it to the field of the automatic coloring of gray comic sketches to solve the problem of the high production cost of color comics and the large consumption of human resources. The method proposed in this paper mainly improves the classical generative adversarial network model in the following aspects: first, the idea of the residual network is added to the builder model and the discriminator model to ensure that the training of the model can develop in the right direction; secondly, add a gradient penalty term to the loss function of the discriminator to accelerate the convergence of the model and speed up the iteration; at the same time, the activation function in the original model is changed to the Mish activation function, so that the information flowing through the network model has higher accuracy and better generalization. This paper trained the model of the Anime Sketch Colorization Pair dataset [1] from Kaggle, and the final experimental results show that the method is practical and feasible.

The traditional human action recognition (HAR) method is based on RGB video. Recently, with the introduction of Microsoft Kinect and other consumer class depth cameras, HAR based on RGB-D (RGB-Depth) has drawn increasing attention from scholars and industry. Compared with the traditional method, the HAR based on RGB-D has high accuracy and strong robustness. In this paper, using a selective ensemble support vector machine to fuse multimodal features for human action recognition is proposed. The algorithm combines the improved HOG feature-based RGB modal data, the depth motion map-based local binary pattern features (DMM-LBP), and the hybrid joint features (HJF)-based joints modal data. Concomitantly, a frame-based selective ensemble support vector machine classification model (SESVM) is proposed, which effectively integrates the selective ensemble strategy with the selection of SVM base classifiers, thus increasing the differences between the base classifiers. The experimental results have demonstrated that the proposed method is simple, fast, and efficient on public datasets in comparison with other action recognition algorithms.

Sparse learning has significant applications in statistics, big data, bioinformatics and machine learning. In big data systems, a large amount of redundant, missing and noisy data cause sparsity, and the rapid changes of information result in uncertainty. Since the traditional sparse learning model is difficult to deal with uncertain data, we propose a Fuzzy Granular Sparse Learning (FGSL) model for identifying antigenic variants of influenza viruses. Firstly, a fuzzy set theory is introduced to measure and granulate the influenza viruses. Some fuzzy granules are induced by a single feature fuzzy granulation. Then, a fuzzy granular vector is constructed from these fuzzy granules, and the fuzzy granular regression is presented. Some constraint norms for granules and granular vectors are proposed, which are two granule norms and four granular vector norms. Therefore, the FGSL model is constructed based on granular regression and constraint norms. The FGSL model includes granular ridge and lasso regressions under different constraint norms. Furthermore, we prove the derivative forms of two granular regression functions, guaranteeing the convergence of the FGSL model. The optimization problem of the FGSL model is discussed and two gradient descent algorithms of the FGSL model are designed. Finally, we employ the FGSL model to serologic data and hemagglutinin sequences for learning antigenicity-associated mutations and inferring antigenic variants. The experimental results confirm some advantages of the FGSL model with fast convergence, low RMSE and strong feature selection ability. We successfully identify antigenic variants of influenza viruses by the FGSL model. (C) 2021 Elsevier B.V. All rights reserved.