Shri Madhwa Vadiraja Institute of Technology & Management

Facial Emotion Recognition (FER) is a computer-based method of recognizing emotions on faces. There is still a shortage on designing FER models that work well on embedded and resource-constrained devices because of the low computational power and memory. The paper introduces a real-time FER scheme that can be used with embedded systems, balancing between the accuracy of recognition and its effectiveness in terms of computational aspects. This approach is suggested using a convolutional neural network alongside transfer learning in order to enhance performance and save training complexity. EfficientNetB0, ResNet50 and MobileNetV2 are three architectures that are tested on the FER-2013 data. Pre-processing of input images is done to store the features of the face and boost model learning. The models are further streamlined to minimize the training time with no impact on classification reliability. Experimental findings reveal that EfficientNetB0 has the highest accuracy of 72.3 % hence it can be applied in real-time embedded applications. ResNet50 offers similar accuracy with more significant computational resources and therefore cannot be used in low-power settings. MobileNetV2 provides a fair balance between accuracy and inference time, which is appropriate in the lightweight edge device. Broadly, the experiment shows that fine-tuned deep learning will be able to effectively realize real-time emotion recognition in an embedded system. The suggested framework can be applied in the intelligence fields of smart surveillance, human-machine interaction, mental health surveillance, and adaptive learning systems.

Communication is an important tool for effective information exchange over people and it ensures messages are clear and as intended, which fosters relationships. Communication between hearing and non-hearing individuals remains limited due to the lack of efficient a utomated sign language recognition systems. The paper outlines a machine learning framework for real-time gesture and sign language detection based on custom dataset. The dataset used includes all 26 letters of the American Sign Language alphabet, each with around 300 samples for training, testing, and validation of the classifiers. $O$ penCV and MediaPipe were used for annotation and accurate landmark feature extraction from the frames captured through the camera interface. These features are then used by a Random Forest Classifier to detect the gestures correctly. The combined approach of random forest along with the MediaPipe model provides a promising precision in the identification o fg estures and, in combination with OpenCV provides robustness in dynamic identification o ver a set of variations in gestures and conditions.

Facial Expression Recognition (FER) plays an important role in affective computing by enabling machines to interpret human emotions through facial cues. However, accurately distinguishing subtle variations between expressions remains a challenging task. Many existing methods rely on a single type of feature descriptor, which may not effectively capture the diverse structural and texture information present in facial images. To address this limitation, this work presents an adaptive feature-level fusion framework that combines three complementary handcrafted descriptors includes Histogram of Oriented Gradients (HoG), Local Binary Patterns (LBP), and a Fast Key Point-Based Descriptor (FKBD). Instead of directly concatenating features, the proposed approach assigns discriminative weights to each descriptor based on class separability, allowing more informative features to contribute more significantly to the final representation while reducing redundancy. The fused feature vector is evaluated using Support Vector Machine (SVM) and k-Nearest Neighbors (kNN) classifiers, with a subject-independent evaluation protocol to ensure unbiased performance assessment. Experiments conducted on the CK+ and FERG-DB datasets demonstrate that the proposed framework achieves high recognition accuracy, reaching 98.74 % with SVM and 95.60 % with k -NN. The results show that combining complementary descriptors with adaptive weighting improves feature discrimination while maintaining low computational complexity, making the approach suitable for real-time and resource-constrained FER applications.

The neutron shielding performance of a 10 × 10 × 10 cm concrete block (density 2.41 g/cm3) was evaluated experimentally and computationally in the mixed-energy neutron field of a 16 Ci 241Am-Be source. Spectrum-integrated neutron capture reaction rates were measured at the incident and transmitted faces of the block using 197Au (n, γ)198Au gold-foil activation and high-purity germanium (HPGe) gamma-ray spectrometry. The measured reaction rates at the front and rear foil positions were 9.25(33) × 10−19 and 5.42(21) × 10−19 atom−1 s−1, respectively. Corresponding Monte Carlo simulations using the OpenMC code with the ENDF/B-VII.1 nuclear data library yielded. 9.88477×10-19 and 5.7117×10-19 atom−1s−1, representing percentage differences of 6.85% and 5.39%. The ratio of the transmitted-to-incident reaction rate agreed to within 1.37% between experiment and simulation, confirming the validity of the transport model for the spectrum-weighted quantity. The neutron transmission factor through the concrete block was determined to be 0.594, corresponding to a ∼ 41% flux reduction relative to the open-beam configuration. Gamma-ray spectrometry of the irradiated concrete identified activation products qualitatively, including 56Mn (846.3 keV, 1809.2 keV), 24Na (1367.2 keV), and naturally occurring 40K (1459.3 keV). Energy-dispersive X-ray spectroscopy (EDX) confirmed the presence of the principal heavy elements (Ca, Si, Fe, Al, K, Mg) in the concrete sample, consistent with the reference composition used in the transport model. The results confirm the reliability of the OpenMC with ENDF/B-VII.1 framework for predicting neutron transport and activation in standard concrete shielding and provide an initial experimental benchmark dataset for future radiation-shielding-related studies at this facility.

This study aims to enhance the understanding of individual-level factors that influence rural women's intentions to engage in sustainable entrepreneurship in India. This study explores the direct impact of perceived capability, social perception, and individual competencies on women's sustainable entrepreneurial intention (SEI). In addition, this study investigated how these variables indirectly affect perceived opportunities. A quantitative methodology was used, with randomly distributed questionnaires among rural women in India. A total of 1250 responses were collected and analyzed using structural equation modeling (SEM) with partial least squares (PLS). The results revealed that perceived capability (beta = 0.103), individual competencies (beta = 0.052), and social perception (beta = 0.226) significantly and positively influenced the SEI. Furthermore, this study indicates that perceived opportunity serves as a mediator in the relationships between perceived capability, social perception, individual competencies, and SEI of women. The proposed model explained 50.9% of the variance in SEI. This study contributes to the existing literature by empirically analyzing the connection between the individual characteristics of women and their intentions to pursue sustainable entrepreneurship, focusing specifically on rural women in India.