Chowdhury Mohammad Masum Refat

Facial expression recognition (FER) is an important component of human--machine interaction, healthcare monitoring, and rehabilitation robotics. Conventional FER systems are predominantly image-based and rely on computer vision techniques, which require high computational resources and are sensitive to environmental conditions such as illumination and occlusion. To address these limitations, this study investigates a wearable sensing approach for FER using stretchable sensors. A comparative evaluation is conducted between a commercially available capacitance-based stretchable sensor and a laboratory-developed resistance-based stretchable sensor fabricated at the International Islamic University Malaysia (IIUM). Time-series data corresponding to four facial expressions---neutral, happy, sad, and disgust---were collected from 30 participants using multiple sensor placements on the face. Statistical features, including mean and standard deviation, were extracted from the sensor signals and used to train and test five supervised machine-learning classifiers: k-nearest neighbor, decision tree, support vector machine, logistic regression, and random forest. Experimental results demonstrate that the commercial sensor achieves higher overall recognition performance, with the random forest classifier yielding an average F1-score of 96%, compared with 90% for the laboratory-developed sensor. The superior performance of the commercial sensor is attributed to its higher sampling rate and greater signal stability. Nevertheless, the laboratory-developed sensor offers advantages in terms of cost, design flexibility, and the ability to measure both stretch and compression. The findings confirm the feasibility of stretchable sensors for facial expression recognition and highlight the potential of laboratory-developed sensors for cost-effective FER systems, particularly in applications such as rehabilitation robotics and human-centered assistive technologies.

Autonomous navigation of cyborg insects in complex environments remains a challenging issue. Cyborg insects, which combine biological organisms with electronic components, offer a unique approach to tackle such challenges. This study presents a biohybrid behavior-based navigation (BIOBBN) system that enables cyborg cockroaches to navigate complex environments autonomously. Two navigation algorithms were developed: reach-avoid navigation for less complex environments and adaptive reach-avoid navigation for more challenging scenarios. This algorithm, especially the second one, leveraged the cockroaches’ natural behaviors, such as wall-following and climbing, to navigate around and over obstacles. Experiments in simulated environments, including sand and rock-covered surfaces, demonstrate the effectiveness of the BIOBBN system in enabling cyborg cockroaches to navigate and reach target locations. The denser second scenario required more time due to increased obstacle avoidance and natural climbing behavior. Overall performance was promising, highlighting the potential of biohybrid navigation for autonomous cyborg insects in navigating complex environments.

This study presents the development of bio‐intelligent cyborg insects (BCI) by utilizing the insect's natural sensory responses and behaviors through a noninvasive feedback control method. While electrical stimulation has been widely used for cyborg insect control, challenges such as habituation and integration with the insects’ natural behaviors remain areas of ongoing research. To address these limitations, this study introduces a control method based on insect visual perception, specifically using their natural aversion to ultraviolet (UV) light. A wearable UV helmet is designed with two UV light‐emitting diodes near the insect's left and right compound eyes. Targeted stimulation of the compound eyes results in directional turning behaviors such as left eye stimulation‐induced right‐turning and right eye stimulation‐induced left‐turning. This study's experiments show that increased UV intensity correspond to larger turning angles, with no evidence of habituation even after repeated trials. Additionally, this method minimizes stimulation by relying on the insect's natural movement patterns, activating the UV stimulus only when the insect is inactive. Compared to electrical stimulation‐based systems, the UV‐based approach significantly reduces the frequency of stimulations while maintaining consistent and reliable control. This finding suggests that noninvasive feedback methods are viable alternatives for guiding cyborg insects in unknown environments.

Facial expressions play a vital role in human communication, especially for individuals with motor impairments who rely on alternative interaction methods. This study presents a deep learning-based approach for real-time emotion classification using stretchable strain sensors integrated into a wearable system. The sensors, fabricated with conductive silver ink on a flexible Tegaderm substrate, detect subtle facial muscle movements. Positioned strategically on the forehead, upper lip, lower lip, and left cheek, these sensors effectively capture emotions such as happiness, neutrality, sadness, and disgust. A data pipeline incorporating Min-Max normalization and SMOTE balancing addresses noise and class imbalances, while dimensionality reduction techniques like PCA and t-SNE enhance data visualization. The system’s classification performance was evaluated using standard machine learning metrics, achieving an overall accuracy of 76.6%, with notable success in distinguishing disgust (86.0% accuracy) and neutrality (81.0% accuracy). This work offers a flexible, cost-effective, and biocompatible solution for emotion recognition, with potential applications in rehabilitation robotics, assistive technologies, and human-computer interaction.

Facial expressions play a vital role in human communication, especially for individuals with motor impairments who rely on alternative interaction methods. This study presents a deep learning-based approach for real-time emotion classification using stretchable strain sensors integrated into a wearable system. The sensors, fabricated with conductive silver ink on a flexible Tegaderm substrate, detect subtle facial muscle movements. Positioned strategically on the forehead, upper lip, lower lip, and left cheek, these sensors effectively capture emotions such as happiness, neutrality, sadness, and disgust. A data pipeline incorporating Min-Max normalization and SMOTE balancing addresses noise and class imbalances, while dimensionality reduction techniques like PCA and t-SNE enhance data visualization. The system's classification performance was evaluated using standard machine learning metrics, achieving an overall accuracy of 76.6\%, with notable success in distinguishing disgust (86.0\% accuracy) and neutrality (81.0\% accuracy). This work offers a flexible, cost-effective, and biocompatible solution for emotion recognition, with potential applications in rehabilitation robotics, assistive technologies, and human-computer interaction.

Wearable sensor technology has shown significant promise in advancing personalized healthcare and rehabilitation. This study presents the implementation of stretchable sensors for facial expression recognition (FER) to control an upper limb rehabilitation robot. Traditional rehabilitation approaches, which rely heavily on therapist-administered exercises, face limitations such as variability in patient engagement and environmental constraints. To address these issues, we developed a low-cost, flexible sensor system capable of detecting subtle facial muscle deformations and translating them into control commands via a Random Forest machine learning algorithm. Our system classifies four emotional states — Neutral, Happy, Sad, and Disgust — with a classification accuracy of 92.4\%. Real-time performance tests on five subjects controlling an elbow rehabilitation robot demonstrated an average accuracy of 75\% for FER. Motor speed adjustments, driven by recognized expressions, dynamically adapt to patient comfort levels, thereby personalizing the rehabilitation process. The stretchable sensor system offers advantages over traditional methods like EMG and computer vision, including greater reliability, low computational demands, and reduced susceptibility to noise.This study highlights the potential of emotion-driven adaptive rehabilitation robots in improving patient outcomes and engagement. Future work will explore refining sensor placement, enhancing machine learning models, and extending the system to other rehabilitation contexts. Integrating such intelligent wearable systems marks a significant step toward more accessible and empathetic tele-rehabilitation solutions.

After a stroke, most patients suffer from various physical movement disorders. Recovery of paralyzed parts can be facilitated through proper rehabilitation exercises. However, many paralyzed patients rely on others for assistance with daily activities and rehabilitation therapy. Consequently, they often do not perform rehabilitation exercises regularly, leading to prolonged recovery periods. In recent years, various rehabilitation robots have become commercially available for home-based exercise environments. However, most of these robots are unable to interpret patients' feedback during rehabilitation exercises. In this paper, a deep learning-based Facial Expression Recognition (FER) system is developed to control a rehabilitation robot. The model is trained using the JAFFE and FER2013 facial expression datasets, and a prototype rehabilitation robot is used to evaluate the system's performance. Experimental hardware tests show that the model achieves an accuracy of 97.01% on the JAFFE dataset and 87% on the FER2013 dataset. In future work, the proposed facial expression model will be tested with an upper limb rehabilitation system to further validate its effectiveness.

Facial expressions play a vital role in human communication, especially for individuals with motor impairments who rely on alternative interaction methods. This study presents a deep learning-based approach for real-time emotion classification using stretchable strain sensors integrated into a wearable system. The sensors, fabricated with conductive silver ink on a flexible Tegaderm substrate, detect subtle facial muscle movements. Positioned strategically on the forehead, upper lip, lower lip, and left cheek, these sensors effectively capture emotions such as happiness, neutrality, sadness, and disgust. A data pipeline incorporating Min-Max normalization and SMOTE balancing addresses noise and class imbalances, while dimensionality reduction techniques like PCA and t-SNE enhance data visualization. The system’s classification performance was evaluated using standard machine learning metrics, achieving an overall accuracy of 76.6%, with notable success in distinguishing disgust (86.0% accuracy) and neutrality (81.0% accuracy). This work offers a flexible, cost-effective, and biocompatible solution for emotion recognition, with potential applications in rehabilitation robotics, assistive technologies, and human-computer interaction.