International Research Journal of Engineering and Technology (IRJET)
e-ISSN: 2395-0056
Volume: 12 Issue: 03 | Mar 2025
p-ISSN: 2395-0072
www.irjet.net
Optimizing Facial Expression Recognition with Deep Hybrid Neural Networks Satyam Kumar1 and Purusharth Agarwal2 1Student, Department of AIML, Manipal University Jaipur, Rajasthan, India
2Student, Department of AIML, Manipal University Jaipur, Rajasthan, India
---------------------------------------------------------------------***--------------------------------------------------------------------systems is challenging. Traditional approaches to FER often Abstract - In Facial Expression Recognition (FER), the
relied on handcrafted features and shallow learning methods. Nevertheless, with the advent of DL, particularly convolutional neural networks (CNNs), significant strides have been made in improving FER accuracy. CNNs are capable of automatically extracting discriminative features from facial images, thus offering promising avenues for advancing FER technology.
landscape has undergone a remarkable evolution in recent years, driven by significant breakthroughs in deep learning(DL), image processing, and cognitive sciences. This study endeavors to push the boundaries of FER precision and efficacy by delving deep into the intricate nuances of facial movement features in static images. The effectiveness of our proposed methodology is substantiated by compelling results shown in Table 1. Notably, Densenet121 + GRU demonstrates exceptional performance, achieving an impressive accuracy of 99.65% on the Facial Emotion Recognition Image Dataset (Table 1). Our methodology capitalizes on the fusion of DL models with RNNs, namely DenseNet121 + GRU, VGG16 + GRU and Xception + GRU, these results underscore the robustness and effectiveness of our approach in accurately discerning facial expressions from static images. Moreover, our findings shed light on the untapped potential of integrating dynamic facial movement features into FER systems. By bridging the gap between static and dynamic characteristics, our methodology holds promise in enhancing the accuracy and reliability of emotion recognition systems in real-world scenarios. In essence, this study contributes to the ongoing advancements in FER. It paves the way for future research endeavors to harness the full spectrum of facial movement features for superior emotion recognition capabilities.
The dataset utilized in this study play a pivotal role in training and evaluating our proposed FER models. The Facial Emotion Recognition Image Dataset (Kaggle) comprises 18,000 images annotated with diverse emotional states, further enriching the training data. This study explores the efficacy of several DL architectures integrated with RNNs for FER tasks. Specifically, we consider DenseNet121 + GRU, VGG16 + GRU and Xception + GRU models. These models are meticulously designed to capture intricate facial expression patterns from static images, thus enabling accurate emotion recognition. A comprehensive methodology detailing these models' architecture and training process is elucidated in Section 4. Notably, our experimentation yields compelling results, with the DenseNet121 + GRU model showcasing the highest accuracy of 99.65% on the Facial Emotion Recognition Image Dataset (Kaggle), underscoring the robustness of our proposed approach. Detailed discussions on these results are provided in Section 5, shedding light on the efficacy of our proposed FER models.
Key Words: Facial Emotion Recognition, Deep Learning, VGG16, Xception, Multimodal Integration, Transfer Learning, Real-world Deployment.
This paper comprehensively investigates DL-based FER, leveraging attention mechanisms to enhance model performance. Through extensive experimentation and analysis, we demonstrate the efficacy of our proposed approach in achieving accuracy levels on diverse FER datasets. The remainder of this paper is organized as follows: Section 2 presents related work in FER, while Section 3 delineates the motivation behind our proposed methodology. Section 4 provides a detailed exposition of our methods, encompassing dataset descriptions, data vectorization techniques, and model architectures. Section 5 presents our experimental results and analyses, followed by concluding remarks and avenues for future research in Section 6.
1. INTRODUCTION Facial expression recognition (FER) is a pivotal domain within AI, garnering increasing attention as it forms the linchpin for effective human-computer interaction (HCL). Emotions, conveyed through facial expressions, are fundamental components of human communication [3]. FER systems aim to decipher these expressions from static images or dynamic video sequences to discern the underlying psychological states of individuals [4]. Furthermore, in fields like counseling psychology, retail sales, social robotics, and e-learning, accurate FER holds profound implications for improving service delivery and user experience. The significance of FER is underscored by its diverse applications across multiple domains. However, the journey towards achieving robust and accurate FER
© 2025, IRJET
|
Impact Factor value: 8.315
|
ISO 9001:2008 Certified Journal
|
Page 997