Vehicle detection is a fundamental task in computer vision, crucial for applications like traffic management and autonomous driving. However, real-world conditions, particularly low-light environments, pose significant challenges. This paper presents a systematic evaluation of vehicle detection performance under different training strategies, using the state-of-the-art models. We compare joint training (daytime and nighttime images) with separate training (daytime or nighttime only) and investigate the impact of Low-Light Image Enhancement (LLIE). Our findings reveal a model-dependent optimal training strategy and the effect of LLIE on detection accuracy. These results highlight the complex interplay between training data, model architecture, and image enhancement, underscoring the need for careful consideration when deploying robust vehicle detection systems in challenging conditions. This research provides crucial guidelines for optimizing training and challenges common assumptions about the relationship between visual quality and detection performance.

Classroom management is a crucial component in ensuring teaching quality. However, traditional methods that rely primarily on manual processes cannot provide continuous assessments in an unobtrusive manner. Recently, previous studies that have explored the use of visual data to support classroom management have yielded promising results. However, the majority focus on recognizing activities at the frame-level or clip-level, which limits their practical application. In this paper, we present a novel method for continuous student activity recognition. We formulate continuous activity recognition as a multiple object tracking (MOT) problem, where each activity instance is treated as a moving object that appears in different regions of the frames over time. The proposed method named STrack4Re (Student Activity Track to Recongize) consists of two steps: a detection step and a tracking step. The main goal of the detection step is to identify the region in the frame where the activity occurs. Once an activity instance is detected, a tracker is initialized to follow this instance over time. In our study, YOLOv5 and OC-SORT have been chosen for object detection and tracking. YOLOv5 has proven effective for static student activity detection, while OC-SORT is a tracking algorithm known for its robustness to occlusions. Therefore, the proposed method could handle occlusion issue in classroom videos. To evaluate the proposed method, we created a new dataset named CStudentAct, which comprises video sequences with appropriate annotations for continuous student activity recognition.

Federated learning (FL) is a security paradigm using low-resource devices to contribute to the training without sharing local data; hence, it is very suitable for practical applications with strict privacy requirements. Although some FL frameworks have been developed to facilitate specific tasks, they are limited to applications that require real and big data, such as human action videos. In this paper, we introduce FlowerAction, a novel framework for human action recognition, built on top of the Flower framework to connect its existing internal functions, including model communication and aggregation algorithms, with our external components (e.g., dataset, models, specific utilities). To the best of our knowledge, this is the first framework of FL for video-based action recognition. Our experimental results on some datasets show that the FlowerAction framework operates properly, helping researchers approach FL with less effort and conduct federated training benchmarks quickly. Furthermore, our analysis results that show performances in terms of top-k accuracy and communication cost of different FL algorithms could give instructive suggestions for selecting and deploying a deep model for FL-based action recognition in the future. Source code is available at https://github.com/quynhkhanh96/flower-action

Pedestrian search based on image or text query often incur high computational costs due to embedding learning. To address this, we propose Att2Search, a method that leverages semantic-level attributes extracted from images, including two main steps: attribute extraction and attribute matching. Extracted attributes are encoded into binary vectors, which are compared using logical XNOR and AND operations. The method supports two types of queries: image-based and attribute-list. Additionally, we use the multilingual Gemini-1.5-Flash LLM to generate Vietnamese text-descriptions from extracted attributes, enabling another matching method. Our experiments are conducted on the 3000VnPersonSearch dataset for Vietnamese person search. Two separate PAR networks that have been trained on the Market-1501 and PA100K datasets are used to extract different semantic-level attributes from the dataset. Att2Search achieves an accuracy of 75.40 % and 68.40% at R@1 when using the XNOR and AND-based matching strategy, respectively, outperforms the state-of-the-art text-based method on the same dataset.

Identifying the botanical origin of honey is essential for ensuring its quality, preventing adulteration, and protecting consumers. Traditional techniques, such as melissopalynology, physicochemical analysis, and PCR, are often labor-intensive, time-consuming, or limited to the detection of only known species, while advanced DNA sequencing remains prohibitively costly. In this study, we aim to develop a deep learning-based approach for identifying pollen grains extracted from honey and captured through microscopic imaging. To achieve this, we first constructed a dataset named VNUA-Pollen52, which consists of microscopic images of pollen grains collected from flowers of plant species cultivated in the surveyed area in Hanoi, Vietnam. Second, we evaluated the classification performance of advanced deep learning models, including MobileNet, YOLOv11, and Vision Transformer, on pollen grain images. To improve performances of these model, we proposed data augmentation and hybrid fusion strategies to improve the identification accuracy of pollen grains extracted from honey. Third, we developed an online platform to support experts in identifying these pollen grains and to gather expert consensus, ensuring accurate determination of the plant species and providing a basis for evaluating the proposed identification strategy. Experimental results on 93 images of pollen grains extracted from honey samples demonstrated the effectiveness of the proposed hybrid fusion strategy, achieving 70.21% accuracy at rank 1 and 92.47% at rank 5. This study demonstrates the capability of recent advances in computer vision to identify pollen grains using their microscopic images, thereby opening up opportunities for the development of automated systems that support plant traceability and quality control of honey.

Federated learning (FL) is a distributed machine learning training paradigm that protects user privacy by training with user data stored on a local device called a client. In realworld FL systems, the number of clients often varies due to both external and internal factors. New clients may join the system right during the training process. Moreover, existing and new clients can have noisy labels with varying levels of noise. Although some works have been proposed to solve the noise problem in FL with the clients that are fixed and determined prior. These methods have not considered the case when new clients join. Therefore, there is a need for a framework that can handle dynamically noisy clients. In this article, we introduce a new framework named FedDC to handle noisy labels in FL systems with dynamic clients. Our FedDC framework is built on top of the 3-stage FedCorr framework which has been designed to work with a fixed number of clients. In our framework, existing noisy clients will be identified through local intrinsic dimensionality (LID) scores. Then to identify new noisy clients, we use a loss threshold combined with the LID scoring technique in the first stage and with only the loss threshold in the second stage. Our experiments on three benchmark datasets that are CIFAR-10 and CIFAR-100 with independent and identically distributed/nonindependent and nonidentically distributed data partition and a real-world noisy dataset, Clothing1M, demonstrate that FedDC helps mitigate the negative impact of new noisy clients and achieves outperformed accuracy compared to FedCorr. Our code is made available at: https://github.com/gttrung/FedDC

In recent years, many efforts have been dedicated to text-based person search, thanks to its potential applications in various domains. However, most of these works focus on person search via queries in English and conduct offline evaluations. Despite some promising results for text-based person search in English, several challenges still prevent its widespread use in practical situations when deployed in minor languages. This article extends person search to the Vietnamese language. In terms of linguistics, English and Vietnamese belong to two different language families. In addition to the difference in vocabulary, these two languages also have opposite word structures and syntactic structures. The contributions of the article are twofold. First, based on the network architecture of the ViTAA model [Wang et al. 2020], a framework for person search through Vietnamese queries has been developed. In this framework, to take into account specific characteristics of the Vietnamese language, the word-tokenizing, Parts of Speech (PoS) tagging techniques of different natural language processing tools, including Underthesea [UndertheseaNLP 2018], SEACoreNLP [Singapore 2021], and PhoNLP [Nguyen and Nguyen 2021], have been investigated in order to extract language elements from Vietnamese descriptions. Our investigation shows that selecting a suitable preprocessing technique can improve person search performance by 1.28% at R@1. When incorporating these preprocessing techniques with the person search model, the best accuracy was achieved with 27.08%, 51.38%, and 63.00% at rank 1, rank 5, and rank 10 respectively. Second, for the first time, an online evaluation of person search through natural language queries was conducted. A web-based application has been developed to serve online evaluation scenarios with different groups of end-users. An extensive evaluation was conducted with 30 subjects and 115 queries. Upon analyzing the experimental results, open issues and suggestions for future improvements in person search were uncovered.

Accurate pollen grain classification is essential to advance ecological research, agricultural monitoring, and public health strategies. However, conventional classification methods are degraded in their performance due to their subjective nature, intensive labor, and high variability in the morphology of pollen grains. This study denotes and addresses these critical issues through an extensive study using advanced deep neural networks. We first introduce a new pollen-grain dataset consisting of 52 species of plants. Although the proposed pollen classifications, such as three state-of-the-art deep learning architectures: YOLOv11, Vision Transformer (ViT), and MobileNet achieve impressive performances, their performances are degraded by the denoted issues. Therefore, the augmentation and late-fusion schema are proposed to handle these issues. Through a comprehensive comparative analysis, we demonstrate the performance of the fusion model, combining ViT and YOLOv11 to achieve greater precision. This study facilitates the reliable inference of floral origins from microscopic imagery.

Classroom activities play a vital role in assessing teaching quality. While numerous studies have investigated activity recognition using classroom visual data, this task remains challenging due to crowded scenes and frequent occlusions. To overcome these challenges, we propose a two-step method for continuous student activity recognition: activity location detection and temporal tracking. For activity detection, we fine-tune two state-of-the-art object detectors to identify relevant regions in each frame. For tracking, we enhance the OC-SORT algorithm by incorporating bounding boxes generated from virtual trajectories and integrating appearance features to reduce tracking loss caused by occlusion. Our proposed method demonstrates robust and effective performance on the CStudentAct dataset, achieving superior tracking accuracy with MOTA improvements ranging from 0.1% to 2.6% across various activities, while also consistently attaining the highest F1-scores in activity recognition—outperforming the baseline by up to 10.9%.

Text line segmentation in historical document images is a crucial task in document analysis and understanding. Historical documents often feature degraded and complex layouts, and they may exhibit distinct writing styles and languages, posing challenges for text line segmentation. Our study focuses on the text line segmentation task for Cham manuscript images. Cham manuscripts, written on Chinese paper and composed in Middle Cham, serve as precious sources to understand the daily lives of the inhabitants of the Champa Kingdom and provide insights into Champa possessions and their management. To address the challenges of text line segmentation in Cham manuscript images, this study proposes an approach that combines deep learning methods with traditional image processing techniques. Our approach takes into account the characteristics of Cham manuscripts to effectively tackle the challenges posed by these documents. Specifically, the procedure begins by leveraging pre-trained deep learning models and includes a fine-tuning step to extract the candidate baseline. Then, various image post-processing algorithms are introduced to further improve segmentation accuracy. Extensive experiments on our Cham manuscripts dataset consisting of 627 manuscripts with approximately 8300 segmented text lines using different evaluation criteria demonstrate the significant superiority of our approach over existing methods. Our method outperforms state-of-the-art performance by a considerable margin.

Person re-identification (Re-ID) is a vital computer vision task focused on matching images of a person of interest as they move across multiple non-overlapping cameras. Thanks to advancements in deep learning models, numerous important milestones have been achieved in the field of person Re-ID. Recent efforts have concentrated on addressing a more realistic scenario where pedestrians are partially occluded. This trend indicates a promising future for the practical implementation of person Re-ID systems. This paper builds upon our previous work, which successfully addressed single-shot person Re-ID using local matching information. For this task, Earth Mover’s Distance (EMD) is employed as a metric to measure similarity between two distributions. To handle multi-shot Re-ID, the proposed framework integrates a feature block, adapting the single-shot methodology to a multi-shot setting. Unlike conventional person Re-ID methods that employ a manually determined images of person, the proposed framework takes a query tracklet as input, which is automatically generated through human detection and tracking steps. To evaluate the proposed method, FAPR dataset (Fully Automated Person ReID) is used. This dataset is one of the few publicly available datasets built specifically for an end-to-end person Re-ID system. Various scenarios are rigorously examined to demonstrate the effectiveness of the proposed framework, especially in challenging conditions with strong occlusion. Across eight experimental scenarios, the proposed method achieves matching rates at rank-1 ranging from 76.3% to 100%. These results underscore the robustness and efficacy of our approach. Our source code is made available at: https://github.com/anhnhust/emd-person-reid.

Traumatic brain injury (TBI) is one of the most prevalent health conditions, with severity assessment serving as an initial step for management, prognosis, and targeted therapy. Existing studies on automated outcome prediction using machine learning (ML) often overlook the importance of TBI features in decision-making and the challenges posed by limited and imbalanced training data. Furthermore, many attempts have focused on quantitatively evaluating ML algorithms without explaining the decisions, making the outcomes difficult to interpret and apply for less-experienced doctors. This study presents a novel supportive tool, named E-TBI (explainable outcome prediction after TBI), designed with a user-friendly web-based interface to assist doctors in outcome prediction after TBI using machine learning. The tool is developed with the capability to visualize rules applied in the decision-making process. At the tool's core is a feature selection and classification module that receives multimodal data from TBI patients (demographic data, clinical data, laboratory test results, and CT findings). It then infers one of four TBI severity levels. This research investigates various machine learning models and feature selection techniques, ultimately identifying the optimal combination of gradient boosting machine and random forest for the task, which we refer to as GBMRF. This method enabled us to identify a small set of essential features, reducing patient testing costs by 35%, while achieving the highest accuracy rates of 88.82% and 89.78% on two datasets (a public TBI dataset and our self-collected dataset, TBI_MH103). Classification modules are available at https://github.com/auverngo110/Traumatic_Brain_Injury_103 .

Physical training for a person with an impaired hand generally requires extensive guidance and supervision from rehabilitation therapists. Currently, the advantages of wearable sensing technologies and the breakthrough performances of deep neural networks offer feasible solutions for developing assistive tools, particularly to automatically monitor the functional use of the patient’s hand. These applications help to report the results of a rehabilitation training, such as the interaction time between hands and objects and the number of interactions during each practice session. This reduces the burden on physiotherapists and occupational therapists and is beneficial for related research, such as rehabilitation assessment and cognitive sciences. However, the lack of datasets, particularly a dedicated dataset for rehabilitation training/exercises that are collected by body-worn sensors, can block these potential advantages. In this paper, we construct such a multimodal dataset and present a comprehensive study of relevant hand tasks using state-of-the-art deep neuronal models. The dataset is collected in a series of rehabilitation exercises of ten patients who are in treatment after injuries or strokes. The constructed dataset includes different wearable modalities such as cameras, accelerometers, and gyroscopes. Table-based exercises for rehabilitation of patients are designed with ball, water bottles, wooden blocks, and cylindrical blocks. In total, 56 exercise videos with more than 54 GB of data are collected. These videos are segmented into 431 sequences that present instances of each exercise at a certain time. To evaluate the performance of recent neuronal networks, we extract 4500 still images to label eight classes including left- and right-hand, and other objects involved in the exercises. In addition, labeling the patient’s hand in a consecutive frames of 40 shots is implemented with approximately 29K images annotated. These labeled datasets are utilized to evaluate three primary tasks related to evaluating physical hand functions: including one- and two-stage neuronal networks for hand detection and identifying left or right patient’s hand; hand tracking with SORT and DeepSORT algorithms. We report a thorough performance analysis of the neuronal network models for the constructed dataset. These results are also thoroughly discussed in terms of technical issues and challenges in developing a robust clinical tool. The dataset and benchmark performances are available to download online: http://mica.edu.vn:50211/rehabhand/.

Artificial Intelligence (AI) plays an increasingly significant role in predicting diseases by leveraging large-scale and complex datasets. AI-based traumatic brain injury (TBI) prediction has traditionally relied on multimodal data, including both structured information like test results and unstructured data such as CT or/and MRI scans. However, one important source of information—physician conclusions in text form at the time of admission and discharge—has largely been underutilized. These textual conclusions offer valuable insights into the patient's condition. This paper proposes a novel approach to classifying TBI severity using physicians' written conclusions. We introduce the viTBI-BERT model, which is built on the ViHealthBERT backbone. Our approach includes pre-processing techniques such as acronym normalization, special character removal, and noun chunking and textual data augmentation. After pre-processing step, the data is passed through the pre-trained ViHealthBERT model, with an additional fully connected layer and a softmax layer for the classification task. The model is tested on three sets of medical notes derived from a self-collected dataset consisting of clinical and CT findings from 503 Vietnamese patients. The viTBI-BERT model achieved a highest sensitivity of 71% for classifying four levels of injury severity. These findings demonstrate the potential of language-based analysis, which, when combined with structured clinical and subclinical data, could further enhance TBI classification outcomes.

Deep learning models typically require large amounts of training data. However, in the medical domain, data annotation is time-consuming and must be performed by domain experts, especially when it involves delineating tumors, such as ovarian masses, in grayscale ultrasound images that are often of low quality. In this study, we propose a diffusion-based data augmentation approach that generates new ultrasound images guided by tumor boundary masks, namely OTADiff. This reduces the manual effort required for detailed region annotation. Ovarian ultra-sound imaging presents two key challenges: i) tumors often resemble surrounding normal tissue, making them difficult to distinguish; ii) ultrasound images typically exhibit signal attenuation, with the lower regions appearing darker than the upper ones. In this paper, we extend the B-Maps approach by additionally addressing the first challenge. Specifically, our method enhances focus on tumor regions to preserve their texture and intensity while downplaying background areas. This results in the generation of tumor-aware synthetic ultrasound images. The generated images are evaluated using the LPIPS metric, and only high-quality samples are retained to enrich the training dataset for detection models. We conduct experiments on two ovarian ultrasound datasets (OUT_2d and OvaTUS) using YOLOv11 and Faster R-CNN detectors. Results show that our tumor-aware augmentation improves image quality and boosts detection mAP by 4% compared to training without augmented data.

Automatic segmentation of ovarian tumors from ultrasound images plays a crucial role in assisting doctors in the diagnosis and decision-making process. However, this task still faces many challenges due to the characteristics of ultrasound images, particularly the unclear boundary of tumors with the background. This paper introduces a method for segmenting ovarian tumors namely called BKSeg-Net using deep learning. Our proposed method is built from a state-of-the-art encoder-decoder architecture (i.e. U-Net) with the integration of an attention gate to capture more correlated features. Besides, we introduce a novel loss function that integrates four distinct loss functions: Jaccard Loss, SSIM Loss, Focal Loss, and Keypoint Loss. While the first three losses are commonly used in the literature, our new component, Keypoint Loss, imposes additional constraints on the tumor boundaries, helping the model learn and predict the regions that better align with the actual tumors. The proposed method is validated on two benchmark datasets, OTU-2D and USOVA3D, demonstrating promising segmentation results with a sensitivity of 90.59% and 97.96% respectively.

The recognition of human activities is a crucial problem in computer vision, aiming to determine the activities of individuals or groups based on a sequence of images or videos. While previous studies have demonstrated feasible results in activity recognition from skeletal joints, challenges arise in conditions with occlusion, complex motions, and inaccurate joint estimations, leading to missing joints that can impact the effectiveness of subsequent recognition steps. This paper proposes a method for recovering missing joints in the estimation of human joints from motion in videos and applying it to activity recognition. With the proposed method, the results of the model’s 3D space estimation with missing joints show great potential, although the model still exhibits some weaknesses in challenging scenarios. For the activity recognition task, the recognition results of the DD-Net model on the Mica-Action3D dataset have improved by up to 25%.

In ovarian cancer diagnosis and treatment, ultrasound is widely used as a screening method thanks to its low cost. Accurately segmenting ovarian tumors from ultrasound images is an important step for further investigation. However, due to the heterogeneity of tumors and the low quality of ultrasound images, the segmentation of ovarian tumors is a challenging task. This paper presents a method for ovarian tumor segmentation from ultrasound images based on Segment Anything Model (SAM). - a transformer based model trained on large-scale datasets. SAM has been proved to be effective for many segmentation tasks. However, SAM traditionally optimizes loss functions based on regions without considering structural similarity constraints between the actual and predicted regions. To enhance model learning, we incorporate IoU, SSIM, and Focal loss functions in SAM model. Furthermore, we utilize two prompts methods: manual prompts and automatic prompts based on the detection results of YOLOv5. Experimental results on OTU2D dataset show that the proposed method outperforms many state of the art methods and the baseline model with 95.12% of Precision when using manual prompts and 91.39% for automatic prompts. Additionally, to evaluate the generalization of the proposed method, we collect a new dataset named OvaTUS dataset and perform cross dataset evaluation.

The number of pollen-bearing honey bees serves as a vital indicator for assessing colony balance and health. Despite its significance, prevailing detection techniques still rely heavily on manual observation and annotation, leading to time-consuming processes that cannot sustain long-term, continuous monitoring efforts. To facilitate automatic beehive monitoring, this study introduces an efficient method for pollen-bearing bee detection. Initially, we furnish a comprehensive dataset, dubbed VnPollenBee, meticulously annotated for pollen-bearing honey bee detection and classification. The dataset comprises 60,826 annotated boxes that delineate both pollen-bearing and non-pollen-bearing bees in 2051 images captured at the entrances of beehives under various environmental conditions. To the best of our knowledge, this represents the first dedicated dataset for pollen-bearing bee detection. The VnPollenBee dataset is publicly accessible to the research community at https://comvis-hust.github.io/datasets/pollenbee.html. Subsequently, we propose the incorporation of diverse techniques into two baseline models, namely YOLOv5 and Faster RCNN, to effectively address the imbalance that arises during the detection of pollen-bearing bees due to their number being typically much lower than the total number of bees present at hive entrances. The experimental results demonstrate that our proposed method outperforms the baseline models on the VnPollenBee dataset, yielding Precision, Recall, and F1 score of 99%, 93%, and 95%, respectively. Specifically, the improvements obtained are 3% and 2% in Recall and F1 score when using YOLOv5, and 3%, 2%, and 2% in Precision, Recall, and F1 score when using Faster RCNN. These findings confirm the potential of our approach to facilitate bee foraging behavior analysis and automated bee monitoring.

Deep learning based automatic classification of benign and malignant ovarian tumors from ovarian ultrasound images brings many significant benefits in women’s health care. Nonetheless, due to privacy concerns associated with medical data, acquiring adequate data for training deep learning models is not straightforward, and sharing data for research purposes is not publicly feasible. Data augmentation helps address this issue by generating new diverse data from existing datasets. In this paper, we explore various data augmentation techniques, ranging from simple methods like geometric transforms and intensity adjustments to advanced techniques using generative AI models such as Deep Convolutional Generative Adversarial Networks (DCGAN). We then introduce a method to automatically assess the quality of generated images. The augmented data is then used to train four state-of-the-art image classification models: VGG16, ResNet34, DenseNet201 and Inception-v3. Experimental results achieved with the OTU2D dataset show an accuracy of 92.54%, increase of 1.92%. On the OvaTUS dataset, the recall reached 93.2%, increase of 5.29%. When combining the OTU2D and OvaTUS datasets, the accuracy was 91.69%, increase of 4.59%. Experimental results show that the generated images significantly improve the classification accuracy for all tested models on three datasets (OTU2D, OvaTUS, and combined OTU2D with OvaTUS).