Authors - Satya Kiranmai Tadepalli, Sujith Kumar Akkanapelli, Sree Harsha Deshamoni, Pranav Bingi Abstract - This paper in detail analyzes how generative AI and encoder-based architectures are drastically changing the realm of video generation with multimodal inputs such as images and text. The application of CNNs, RNNs, and Transformers so neatly serves to encode divergent modalities that blend into the seamless synthesis of realistic video sequences. It is based on the up-and-coming fields of generative models like GANs and VAEs, in bridging from static images to video generation. However, this represents a significant leap forward in the technology of video creation. It also goes into great detail on the complexities of multimodal input, working to balance coherence over time as well as semantic alignment of what's being produced. In the above-described context, it can now be realized how the role of encoders translates visual and textual information into actionable representations for generating video. What follows is a survey on recent progress in adopting Generative AI and multimodal encoders, discussions on what challenges are encountered today, and possible future directions that ultimately lay emphasis on their potential to assist video-related tasks and change the multimedia and AI communities.
Authors - Vathsal Tammewar, Bharat Sharma, Dharti Sorte, R. Sreemathy Abstract - Accelerated facial aging using GANs has been the key interest area in generative modeling and facial analysis fields, which offers significant breakthrough in age progression and regression solutions. This survey conducted an extensive review on techniques of GAN- based approaches for accelerated facial aging, emphasizing highly realistic and controllable aging transformations. Many of these methods applied in forensic investigations, entertainment industries, or age-invariant facial recognition systems, which are vivid demonstrations of the versatility and practical relevance of such methods. While such recent breakthroughs hold great promises, several issues remain; namely high-fidelity transformations to preserve important facial details do not fully diminish biases due to imbalanced datasets, and temporal consistency when age progressions or regressions consist of sequential ages is also critical. Computational efficiency and real-time applicability are still the most critical areas of focus. This paper probes into the strengths, limitations, and open challenges of existing approaches, while emphasizing the importance of innovations such as improved loss functions, diverse and representative training datasets, and hybrid architectures. Thus, this survey contributes to synthesizing current progress and outlining future research directions for advancing the field of GAN-based facial aging technologies.
Authors - Khushi Mantri, Abhishek Masne, Shruti Patil, Girish Mundada Abstract - In medical diagnostics, identifying bone fractures is a crucial task that is traditionally dependent on radiologists deciphering X-ray pictures. However, human factors like experience or exhaustion can occasionally cause delays or inaccuracies in diagnosis. The construction of an automated system for bone fracture identification utilizing Convolutional Neural Networks (CNN), a deep learning method that performs especially well in picture processing, is examined in this research. With the use of a tagged dataset of X-ray pictures, the suggested method can efficiently and accurately detect fractures. Prior to feature extraction using CNN layers which are trained to distinguish between fractured and non-fractured bones the images are pre-processed to improve clarity. In order to assist medical practitioners in making prompt, correct judgments, the final classification attempts to increase diagnostic accuracy while decreasing the amount of time needed for analysis. The potential of incorporating machine learning into healthcare to lower diagnostic errors and enhance patient outcomes is also discussed in this overview paper, which includes examines recent developments in CNN-based medical picture categorization.
Authors - Shradha Jain, Sneha Suman, Insha Khan, Ashwani Kumar, Surbhi Sharma Abstract - With continual advancements in deep learning, the potential misuse of deep fake is increasing and its detection is in a major scope of work. A model is trained to recognize the patterns in input data, deep fake recognize those patterns in a fabricated way. Sometimes a small, intentional change is added in the data points, these changes are undetectable to humans and confuse the learning model. Those changes are called adversarial perturbations. Compressive adversarial perturbations aim to make those changes even smaller and harder to detect. Authors explore a sophisticated framework - ComPAD (Compressive Adversarial Perturbations and Detection) which is used to detect adversarial attacks. This paper explores the strategies, and provides comparative analysis of methods used by different researchers. Various datasets including UADFV, DeepfakeTIMIT, LFW, FF++, and Deeperforensics are evaluated to achieve the highest metrics. Methods based on convolution neural networks, particle swarm optimization, genetic algorithm and D4 (Disjoint Diffusion Deep Face Detection) are used for detection. Authors also discuss the challenges such as generalization of models across the new data, the continuous evolution of adversarial perturbations that leads to consistent attacks, and the scalability issues for the real time deep fake. Concluding that models can significantly improve the accuracy, robustness and generalization.
Authors - Krisha Zalaria, Jaitej Singh, Priyanka Patel Abstract - The ubiquitous use of mobile phones in modern society has sparked increasing concern in environments where their usage is restricted, such as hospitals, schools, religious sites, and hazardous zones. Mobile phones, although integral to daily life, pose risks such as privacy breaches, interference with sensitive equipment, and even serious safety hazards. In response, this paper investigates the efficacy of various state-of-the-art object detection models for real-time mobile phone detection in restricted areas. We benchmarked YOLOv8, YOLOv9, EfficientDet, Faster R-CNN, and Mask R-CNN to identify optimal solutions balancing speed, accuracy, and adaptability. This study introduces a two-class detection framework to distinguish between individuals texting or talking on the phone, catering to differing levels of restriction. Evaluations using a customized, diverse dataset reveal YOLOv8 and YOLOv9 as superior, achieving high precision and recall, thus positioning these models as effective solutions for scalable, real-time surveillance systems in sensitive environments. Our research contributes significant insights into mobile phone detection, paving the way for enhanced safety and privacy in restricted zones.
Authors - Aniket Gupta, Chris Dsouza, Sarah Pradhan, Amiya Kumar Tripathy, Phiroj Shaikh Abstract - This paper is based on the development of voice chatbots and their configuration to make sure that e-commerce websites are in compliance with all the customer care requirements. The authors talk about the introduction of natural language processing in an e-commerce company and provide a review of recent developments in that area. The research specifically focuses on natural language processing techniques, steps involved in developing a chatbot, problems encountered during design, and functions and benefits of voice-based chatbot in e-commerce. This A study emphasizes chatbots as tools to support customer service systems. Keywords: Machine Learning, Natural Language Processing, Data Analysis, Customer support system, and CHATBOT.
Authors - Siddharth Lalwani, Abhiram Joshi, Atharva Jagdale, M.V.Munot, R. C. Jaiswal Abstract - Empathetic response generation is a rapidly evolving field focused on developing AI systems capable of recognizing, understanding, and responding to human emotions in a meaningful way. This paper investigates the integration of the Big Five OCEAN personality model with generative AI to generate emotionally relevant, personalized responses tailored to individual users' personality traits. The Big Five model categorizes individuals into five core personality dimensions—Openness, Conscientiousness, Extraversion, Agreeableness, and Neuroticism. By combining this model with advanced generative AI techniques, the system can deliver empathetic responses aligned with users' emotional states and personality profiles. Through the use of various machine learning algorithms, the study demonstrates that incorporating personality traits significantly improves the quality, accuracy, and emotional resonance of AI-generated responses, leading to more effective human-AI interactions.
Authors - Shaga Anoosha, B Seetharamulu Abstract - Brain tumor segmentation is a critical task in medical imaging, essential for accurate diagnosis and treatment planning. Recent advancements in federated learning (FL) and deep learning (DL) offer promising solutions to the challenges posed by traditional centralized learning methods, particularly regarding data privacy and security. This review paper delves into state-of-the-art approaches that FL and DL to enhance brain tumor segmentation. Each institution trains a deep learning model, typically a Convolutional Neural Network (CNN) or a specialized architectures like U-Net on its local dataset. U-Net, particularly effective for image segmentation tasks, consists of an encoder that extracts hierarchical features from MRI scans and a decoder that reconstructs the segmented output, creating a segmentation map outlining tumor boundaries. Instead of sharing raw MRI scans, federated learning allows each institution to share model updates with a central server. The central server aggregates the updates from all participating institutions to create a global model using Federated Averaging, which averages the weights of the local models. The updated global model is then sent back to each institution, which continues training on their local data using this improved model. This iterative process ensures high accuracy, robustness, and privacy preservation, making it a promising approach for collaborative brain tumor detection and segmentation. By combining the strengths of federated learning and deep learning, these state-of-the-art methodologies provide a powerful solution to the challenges posed by traditional centralized models. This integration not only improves segmentation performance but also ensures that sensitive patient data remains secure. As advancements in this field progress, the collaborative use of these state-of-the-art techniques is poised to significantly enhance diagnostic accuracy and improve patient outcomes in medical imaging.
Authors - Adarsh Singh Jadon, Rohit Agrawal, Aditya A. Shastri Abstract - This study investigated the efficacy of various deep learning models in performing sentiment analysis on code-mixed Hinglish text, a hybrid language widely used in digital communication. Hinglish presents unique challenges due to its informal nature, frequent code-switching, and complex linguistic structure. This research leverages datasets from the SemEval-2020 Task 9 competition and employs models such as RNN (LSTM), BERT-LSTM, CNN, and a proposed Hybrid LSTM-GRU with 1D-CNN model. Combining the strengths of LSTM and GRU units along with 1D-CNN, demonstrated superior performance with an accuracy of 93.21%, precision of 93.57%, and recall of 93.02%, along with Sensitivity & Specificity of 93.62% and 93.24% respectively. It also achieved F1 Score of 93.44%. We also evaluated the model on some other parameters, such as PPV, PNV, RPV, and RNV. This model outperformed existing approaches, including the HF-CSA model from the SemEval-2020 dataset, which achieved an accuracy of 76.18%.
Authors - Chetana Shravage, Shubhangi Vairagar, Priya Metri, Akanksha Madhukar Pawar, Bhagyashri Dhananjay Dhande, Siddhi Vaibhav Firodiya, Tanmay Pramod Kale Abstract - Emotion Recognition has gained significant popularity, driven by its wide range of applications. Emotion recognition methods use various human cues such as use of speech, facial expressions, body postures or body gestures. The methods built for emotion recognition use combinations of different human cues together for better accuracy in results. This paper explores different methods which use different human cues.
