HYBRID APPROACH FOR CRICKET HIGHLIGHT GENERATION BASED ON PLAYER-SPECIFIC ANALYSIS

Abstract

Highlight generation involves extracting the most engaging clips from a sports video. In the context of video summarization, the entire video is condensed into a shorter format, retaining the most critical content. For example, a complete match video in cricket includes various actions such as fours, sixes, and wickets. Highlights, on the other hand, still these significant events—fours, sixes, and wickets—into a cohesive and essential highlights package. We use recorded cricket videos and player images from publicly accessible online sources to conduct our study. Data pretreatment is done first, and then cleaned data is ready for deep learning model training. CNN and VGG-16 are the two deep-learning models that we trained. Following the creation of the models, standard assessment metrics are used to compare the two models. We trained the model for 70 epochs with SGD optimizer, and categorical cross-entropy loss function to optimize the model parameters. We trained two models for the task of cricket player recognition and highlight generation: a CNN model for frame extraction and a pre-trained VGG-16 model for feature extraction. The results suggest that both models are adequate for the task of cricket player recognition and highlight generation. The proposed model achieved 96% accuracy in the testing phase of the study. However, the VGG-16 model achieved higher accuracy than the CNN model, indicating that the pre-trained model is more effective at extracting relevant features from the frames. The ROC curves also suggest that both models have good discrimination ability, which is essential for generating accurate player-specific highlights.