A Study on Video Classification Based on Deep Learning
Zhangmeizhi Li · 2024
This research delves into the application of deep learning methodologies for video classification, with a specific focus on sports action recognition. Utilizing the A1 sports action dataset, the study partitions the data into 75% for training and 25% for testing, aiming to assess the efficacy of a novel proposed method against existing classification techniques. These techniques range from traditional handcrafted feature approaches to advanced deep learning models, including SGSH, snippets, and two-stream LSTM, with the proposed method demonstrating superior performance through meticulous evaluation. Through an experimental framework that incorporates the ResNet50 network and a temporal frame sampling strategy (T = 16), the research meticulously adjusts video frame preprocessing for optimal deep learning model training and testing. Pretraining on the ImageNet dataset and employing a dynamic learning rate adjustment strategy, the study ensures a robust evaluation of the proposed method's classification accuracy. Notably, the proposed method achieved a remarkable mean accuracy of 99.6%, outperforming the two-stream LSTM model by 0.7% and showcasing its exceptional capability in accurately classifying a wide range of sports actions. The findings from this study highlight the significant potential of deep learning in enhancing video classification accuracy, especially in the context of sports action recognition. By effectively capturing both short-term and long-term motion dynamics through innovative network architectures and processing strategies, the proposed method sets a new benchmark in the field. This research contributes valuable insights into the evolving landscape of video classification, underscoring the importance of deep learning techniques in addressing complex classification challenges and paving the way for future advancements in the domain.