Recognition Learning Research Articles

Balanced clustering contrastive learning for long-tailed visual recognition

Balanced clustering contrastive learning for long-tailed visual recognition

Relaxed Relational Knowledge Distillation for Inter–Intra Class Learning in Human Activity Recognition

Relaxed Relational Knowledge Distillation for Inter–Intra Class Learning in Human Activity Recognition

Advancing precision agriculture with deep learning enhanced SIS-YOLOv8 for Solanaceae crop monitoring

IntroductionPotatoes and tomatoes are important Solanaceae crops that require effective disease monitoring for optimal agricultural production. Traditional disease monitoring methods rely on manual visual inspection, which is inefficient and prone to subjective bias. The application of deep learning in image recognition has led to object detection models such as YOLO (You Only Look Once), which have shown high efficiency in disease identification. However, complex climatic conditions in real agricultural environments challenge model robustness, and current mainstream models struggle with accurate recognition of the same diseases across different plant species.MethodsThis paper proposes the SIS-YOLOv8 model, which enhances adaptability to complex agricultural climates by improving the YOLOv8 network structure. The research introduces three key modules: 1) a Fusion-Inception Conv module to improve feature extraction against complex backgrounds like rain and haze; 2) a C2f-SIS module incorporating Style Randomization to enhance generalization ability for different crop diseases and extract more detailed disease features; and 3) an SPPF-IS module to boost model robustness through feature fusion. To reduce the model’s parameter size, this study employs the Dep Graph pruning method, significantly decreasing parameter volume by 19.9% and computational load while maintaining accuracy.ResultsExperimental results show that the SIS-YOLOv8 model outperforms the original YOLOv8n model in disease detection tasks for potatoes and tomatoes, with improvements of 8.2% in accuracy, 4% in recall rate, 5.9% in mAP50, and 6.3% in mAP50-95.DiscussionThrough these network structure optimizations, the SIS-YOLOv8 model demonstrates enhanced adaptability to complex agricultural environments, offering an effective solution for automatic crop disease detection. By improving model efficiency and robustness, our approach not only advances agricultural disease monitoring but also contributes to the broader adoption of AI-driven solutions for sustainable crop management in diverse climates.

The use of deep learning in intelligent athlete motion recognition: Integrating biological mechanisms

This work explores the effective application of deep learning for recognizing athletes’ movements, aiming to enhance precision in competitive sports. Traditional motion analysis methods primarily rely on manual observation, which can introduce subjective bias and limit accuracy. To address these limitations, we propose an automated method based on deep learning for recognizing and classifying athletes’ technical movements while evaluating their performance. A hybrid model, combining Convolutional Neural Networks (CNN) and Long Short-Term Memory (LSTM) networks, is utilized to extract key frames from video data. The CNN is responsible for feature extraction, capturing the intricate details of movement, while the LSTM captures the temporal sequence characteristics, providing context to the actions. To further strengthen our approach, we delve into the biological mechanisms underlying athletic movements. Understanding the biomechanics of motion—such as joint angles, muscle activation patterns, and energy expenditure—can enhance the accuracy of deep learning models. By integrating these biological insights into our model, we improve the recognition process, allowing for a more nuanced understanding of how movements impact performance. Through experiments, we demonstrate that the model achieves high accuracy across multiple benchmark datasets (UCF-101, HMDB-51, Kinetics-400, and Sports-1M), with a particularly high accuracy of 93.5% on the UCF-101 dataset. These results indicate that the proposed method is both accurate and reliable, making it suitable for athlete training and competition analysis. The findings of this research have significant implications for sports science, training evaluation, and injury prevention. By providing coaches and athletes with precise feedback based on deep learning analysis, we can facilitate targeted training interventions that enhance performance while reducing injury risks. This work aims to offer a powerful tool for athletes, coaches, and researchers, contributing to the advancement of competitive sports through a deeper understanding of movement dynamics and their biological underpinnings.

Human vs machine learning in face recognition: a case study from the travel industry

This research was conducted to help answer whether a machine learning simulation can replace the human ability to recognize human faces, especially under challenges under travel industry requirements. The human ability to recognize faces was evaluated using a series of questions in a survey. The questions challenged the human respondents to recognize faces under similar looks, with hair and makeup disguises, only part of the facial area, and under dark lighting conditions. At the same time, a histogram of oriented gradient (HoG) combined with a support vector machine (SVM) was built for machine learning simulations. The machine learning was evaluated using two datasets, i.e., the Extended Yale B (EYB) Face dataset for challenge under dark lighting conditions and The Extended Makeup Face Dataset (EMFD) for challenge using face with makeup disguise. The results showed that machine learning simulation of the face recognition system yielded accuracy as high as 95.4% under dark lighting conditions and 70.8% under facial makeup disguise. On the contrary, only 48% of respondents accurately recognized human faces in dark lighting. The number was increased to 94-96% when the face images were adjusted first with the contrast adjustment method. However, only 36-37% of respondents accurately recognized human faces under face makeup disguise.

Using voice recognition and machine learning techniques for detecting patient-reported outcomes from conversational voice in palliative care patients.

Patient-reported outcome measures (PROMs) are increasingly used in palliative care to evaluate patients' symptoms and conditions. Healthcare providers often collect PROMs through conversations. However, the manual entry of these data into electronic medical records can be burdensome for healthcare providers. Voice recognition technology has been explored as a potential solution for alleviating this burden. However, research on voice recognition technology for palliative care is lacking. This study aimed to verify the use of voice recognition and machine learning to automatically evaluate PROMs using clinical conversation voice data. We recruited 100 home-based palliative care patients from February to May 2023, conducted interviews using the Integrated Palliative Care Outcome Scale (IPOS), and transcribed their voice data using an existing voice recognition tool. We calculated the recognition rate and developed a machine learning model for symptom detection. Model performance was primarily evaluated using the F1 score, harmonic mean of the model's positive predictive value, and recall. The mean age of the patients was 80.6 years (SD, 10.8 years), and 34.0% were men. Thirteen patients had cancer, and 87 did not. The patient voice recognition rate of 55.6% (SD, 12.1%) was significantly lower than the overall recognition rate of 76.1% (SD, 6.4%). The F1 scores for the five total symptoms ranged from 0.31 to 0.46. Although further improvements are necessary to enhance our model's performance, this study provides valuable insights into voice recognition and machine learning in clinical settings. We expect our findings will reduce the burden of recording PROMs on healthcare providers, increasing the wider use of PROMs.

Feature selection through adaptive sparse learning for scene recognition

Scene recognition is an important and challenging task in the field of computer vision. Current research typically focuses on local features in scene images by utilizing pretrained convolutional neural networks (CNN) to obtain a feature dictionary. However, these local features often contain similar features to other scene categories, leading to feature confusion and subsequently low accuracy in scene recognition. Therefore, focusing on local features for scene recognition is controversial. In contrast to existing works, we consider extracting common features within the same category from scene images and using them as discriminative features between different categories. We propose a scene recognition method based on adaptive sparse learning feature selection. This method leverages sparse learning to distinguish the contribution of each feature in the deep features to the classification task, aiming to construct a salient feature dictionary that describes the importance of features in scene images. Subsequently, an adaptive weight optimization method is employed through alternating updates to automatically adjust feature weights, enabling the selection of common features within the same scene category from the compact dictionary. Experimental results on multiple benchmark scene recognition datasets demonstrate that our proposed method is superior to state-of-the-art methods.

GCD-JFSE: Graph-based class-domain knowledge joint feature selection and ensemble learning for EEG-based emotion recognition

Feature selection has demonstrated strong performance in emotion recognition using intrasubject electroencephalography (EEG) data. However, it faces challenges due to individual differences and the nonstationarity of EEG signals in cross-subject and cross-session emotion recognition. Currently, research on incorporating domain information into feature selection for cross-domain (subject or session) emotion recognition remains limited. To address this issue, we propose a graph-based class-domain knowledge joint feature selection and ensemble learning approach. Firstly, an undirected, fully connected weighted graph is constructed to capture the relationship between features. Then, some metrics such as domain scatter, domain correlation, and domain standard deviation are introduced to guide feature selection. Subsequently, soft voting ensemble learning is employed to enhance recognition performance. To validate the effectiveness of our method, we conduct experiments on public datasets (SEED, SEED_IV, DREAMER), achieving accuracies of 78.67% on SEED, 58.98% on SEED_IV, 61.11% of valence and 72.46% of arousal on DREAMER in a cross-subject scenario. In the cross-session scenario, we obtain 87.11% on SEED and 60.74% on SEED_IV. The proposed method outperforms state-of-the-art approaches. This study not only expands the application of feature selection in emotion recognition but also provides a potential strategy to enhance the performance of real-world EEG-based emotion recognition applications.

SDR-GNN: Spectral Domain Reconstruction Graph Neural Network for incomplete multimodal learning in conversational emotion recognition

SDR-GNN: Spectral Domain Reconstruction Graph Neural Network for incomplete multimodal learning in conversational emotion recognition

IT and Management Strategies for Language Education: Lessons from the Digitalization of Education Activities

AbstractThe Education 4.0 plan had initiated major changes to the education sector’s workflow and teaching method. In Indonesia, this sudden transformation is perceived as both an opportunity and a drawback, especially in language learning. This study investigates how technology can be integrated in language education, with the purpose of designing a strategy that addresses pedagogical problems, enhances the implementation of technology, and increases learning effectiveness among students. The methods used in this research are Mixed-method, Comparative analysis, and Framework development. The result shows that while the shift to online learning presented challenges for both teachers and students, such as technological proficiency and maintaining engagement, it also opened up opportunities to utilize innovative technologies like Automatic Speech Recognition (ASR) and game-based learning to enhance language acquisition. In conclusion, adoption of Information Technology (IT) in the language education sector has its own challenges and opportunity, comprehensive understanding of how IT influences language education during and after pandemic to better guide technological innovation to the betterment of the language education sector.

초등 실과에서 효율적인 소프트웨어와 로봇 및 인공지능 교육을 위한 플랫폼의 개발

This study aimed to develop an educational platform to efficiently implement software, robotics, and artificial intelligence education in the revised 2022 elementary practical arts education curriculum. The results of the study are as follows. The revised 2015 curriculum posed various methodological difficulties that hindered efficient software and robotics instruction. First, there were difficulties in pairing block-based programming tools (software) with robots, the hardware that executes coding content. Second, teachers found monitoring each student's coding progress in real-time difficult, and answering individual questions required additional time, making personalized guidance difficult. Third, it took considerable time to share correctly coded programs with students for comparison and modification or to explain the correct coding methods. To address these challenges, this study developed a platform to efficiently facilitate software, robot, and coding education for artificial intelligence. The platform allows for easy pairing between programming tools and hardware and allows teachers and students to share coding content in real time. In addition, students can copy correctly coded programs to their own platforms in real time and modify or execute them. This platform is compatible with not only computers but also tablets and mobile phones, with capabilities such as voice recognition, image recognition, and machine learning. Image recognition functionality includes blocks for face, hand, and character recognition.

Development of "Target Detection and Identification System" for Combat Military Vehicles via Artificial Intelligence and Decision Support Matrix

Modern military intelligence systems are improving day by day. In the context of ground forces, combat vehicle-type classification technology serves as a valuable supporting asset. Understanding the capabilities and limitations of different combat vehicle types is crucial for developing effective tactics and strategies. The implementation of artificial intelligence-supported military vehicle type classification technology has raised increasing concerns as image processing, pattern recognition, and deep learning have advanced. You Only Look Once (YOLO) has demonstrated numerous considerable advantages in object detection and image classification. This method accelerates target detection because it can predict objects in real-time. The high accuracy of detection and assessment assists operational personnel in the field. The YOLO prediction method produces accurate results with minimal background errors and facilitates the understanding of generalized object representations. This paper utilizes YOLO to demonstrate combat vehicle type detection, specifically employing the YOLOv8m model, which is well-suited for mobile deployments. This study focuses on a process that begins with detecting targets using images obtained through electro-optical systems, aimed at supporting the activities of identifying and defining targets, which are essential components of target management systems. The process involves developing a Target Detection and Identification (TDI) system based on deep learning, incorporating steps such as pre-processing, segmentation, feature extraction, classification, additional data extraction, and decision support. This system is designed to interpret the results obtained from these processes and provide actionable recommendations. Furthermore, it assists in addressing the weapon target assignment problem in the land environment.

Perancangan Mesin Klasifikasi Menggunakan Particle Swarm Optimization

Designing an effective classification engine is very important in various pattern recognition and machine learning applications. In this research, the Particle Swarm Optimization (PSO) algorithm is applied for the development of classification engines on various datasets. PSO is a population-based optimization method inspired by the behavior of flocks of birds or fish, which is effectively used to find optimal solutions in large search spaces. This research aims to develop a classification model by using Particle Swarm Optimization (PSO) as a training element to determine weights and biases. To test the performance on several different datasets, namely on a dummy multi-class dataset, Sasak Aksara image dataset, and the well-known Iris dataset. In the Sasak Aksara data, Discrete Cosine Trasnform (DCT) is used as feature extraction with the aim of reducing computation time. The results show that PSO can be used in the implementation of several datasets used, in the classification of dummy data, iris data, and Sasak Aksara image data. The model achieved 100% accuracy, precision, recall, and F1-Score on dummy data and iris data. However, on the Sasak Aksara image dataset, the performance of the model decreased with accuracy only reaching 65%, precision 50%, recall 32%, and F1-Score 39%. This research contributes in demonstrating the effectiveness of PSO in optimizing Perceptron models on simpler datasets and highlights the need for further development to handle more complex datasets.

AI-Driven Econometric Models for Legal Issues

Artificial intelligence (AI) is reshaping the landscape of econometric modeling, offering innovative tools to address complex legal issues involving predictive analysis, resource allocation, and policy evaluation. This research explores the application of AI-driven econometric models to legal challenges, focusing on areas such as contract enforcement, intellectual property disputes, and regulatory compliance. By integrating machine learning with traditional econometric techniques, these models enhance the precision and adaptability of legal forecasts and decision-making processes. Key methodologies include the use of natural language processing (NLP) for legal text analysis, deep learning for pattern recognition in case law, and game theory to evaluate strategic interactions in legal contexts. The study highlights the potential of AI to improve efficiency, reduce bias, and facilitate equitable outcomes in legal systems. Challenges such as data privacy, interpretability, and ethical considerations are also addressed. By bridging AI and econometrics, this research aims to provide a robust framework for advancing legal analytics, contributing to more informed policy-making and judicial processes.

Graph-Enhanced Low-Resource ECG Representation Learning for Emotion Recognition Based on Wearable Internet of Things

Graph-Enhanced Low-Resource ECG Representation Learning for Emotion Recognition Based on Wearable Internet of Things

Enhanced multi-branch learning for long-tailed image recognition

Enhanced multi-branch learning for long-tailed image recognition

Self-supervised representation learning for robust fine-grained human hand action recognition in industrial assembly lines

Humans are still indispensable on industrial assembly lines, but in the event of an error, they need support from intelligent systems. In addition to the objects to be observed, it is equally important to understand the fine-grained hand movements of a human to be able to track the entire process. However, these deep-learning-based hand action recognition methods are very label intensive, which cannot be offered by all industrial companies due to the associated costs. This work therefore presents a self-supervised learning approach for industrial assembly processes that allows a spatio-temporal transformer architecture to be pre-trained on a variety of information from real-world video footage of daily life. Subsequently, this deep learning model is adapted to the industrial assembly task at hand using only a few labels. Well-known real-world datasets best suited for representation learning of such hand actions in a regression tasks are outlined and to what extent they optimize the subsequent supervised trained classification task. This subsequent fine-tuning is supplemented by concept drift detection, which makes the resulting productively employed models more robust against concept drift and future changing assembly movements.

A Novel Technique for Handwritten Signature Recognition

Handwritten signature recognition (HSR) is a critical component of biometric systems, widely used for securing financial transactions and identity verification. However, the variability of handwritten signatures, influenced by individual writing styles, inconsistencies, and environmental factors, presents significant challenges for recognition systems. Despite these obstacles, signatures remain a reliable and popular biometric trait. This paper introduces a novel deep learning approach utilizing a convolutional neural network (CNN) architecture specifically designed for HSR. The proposed method was validated using two prominent datasets, MCYT-75 and GPDS-300, with detailed descriptions of the CNN structure. Experiments, conducted on a personal computer equipped with an NVIDIA Quadro M1200 GPU, an Intel i7 processor, and 32 GB of RAM, demonstrated the model’s exceptional performance, achieving validation accuracies of 99.60% on the MCYT-75 dataset and 99.80% on the GPDS-300 dataset. These results reflect the model’s robustness and minimized error rates, outperforming existing techniques and underscoring the effectiveness of deep learning for signature recognition. This study highlights the proposed model's potential for real-world applications and paves the way for further advancements in biometric authentication technologies.

Deep Learning and Visual SLAM for Autonomous Navigation of UAVs: Status, Challenges, and Future Perspectives

This paper discusses the application and integration of deep learning and visual SLAM technology in UAV autonomous navigation. With the wide application of UAVs in the fields of transportation, agriculture, military and environmental monitoring, improving the autonomous navigation capabilities of UAVs has become an important demand. This paper first analyses the advantages and disadvantages of deep learning and visual SLAM. Then, this paper emphasizes the core role of deep learning in feature extraction, target recognition and path planning. The key role of visual SLAM in real-time localization and environment mapping is highlighted as well. This paper discusses the combined application of these two technologies and demonstrates how to enhance the stability and accuracy of visual SLAM systems through deep learning in complex dynamic environments. Although the current technology still faces high computational requirements and real-time challenges, this paper proposes future research direction to promote the continued development of UAV autonomous navigation technology.

Deep learning utilization in Sundanese script recognition for cultural preservation

This study addresses the challenge of preserving the Sundanese script, a traditional writing system of the Sundanese community in Indonesia, which is at risk of being forgotten due to technological advancements. To tackle this problem, we propose a deep learning approach using the YOLOv8 model for the automatic recognition of Sundanese characters. Our methodology includes creating a comprehensive dataset, applying augmentation techniques, and annotating the characters. The trained model achieved a precision of 95% after 150 epochs, demonstrating its effectiveness in recognizing Sundanese characters. While some variability in accuracy was observed for certain characters and real-time applications, the results indicate the feasibility and promise of using deep learning for Sundanese script recognition. This research highlights the potential of technological solutions to digitize and preserve the Sundanese script, ensuring its continued legacy and accessibility for future generations. Thus, we contribute to cultural preservation by providing a method to safeguard the Sundanese script against obsolescence.