Sports Genre Detection Based On Deep Learning

Along with the rapid development of internet and computer techniques, a large number of sports video information is produced. Automatically Annotate and positioning highlight clips in videos has become one of the attractive and challenging researches in computer vision areas. Deep learning achieved amazing performance in pattern recognition, and deep CNN model has been applied to various tasks, including image classification, object detection, face detection and recognition. Deep Neural Networks boosts the performance of computer vision to a new height.This paper aims to optimize a large CNN model on a small scale dataset, with the help of a large auxiliary dataset. Deep architecture performs well on large dataset, but if suffers from over fitting on a small dataset even with dropout and data augmentation. Transferring big structure trained on a large dataset to a new irrelevant dataset is meaningful. We proposed two algorithms to implement the transfer:Fine-tune a deep CNN model trained on a large auxiliary dataset by replacing the classifiers of the CNN model of a new one and remaining the feature extractor. We treat the stacked convolutional layers and pooling layers as feature extractor, and the fully-connected layers and SoftMax layer as a MLP classifier. Additionally, we compared the performance of CNN feature and handcraft features, such as SIFT. Thus, feature vectors extracted from the CNN model and SIFT-BOW framework are fed to a linear SVM classifier.Combine the computer vision and text semantics. The probability distribution on the pre-trained dataset is transform to the new data through lexical semantic distance of the concepts of the two datasets. We tried two algorithms to form the mapping matrix:Normalized Google distance and WordNet-based measures.Deep Neural Networks is highly abstract and it’s difficult for us to analyze the features of the higher layers directly. We introduced deconvolutional network to map the features of the higher layers to pixel space (RGB space), thus we could analysis the features using our eyes. It’s important for us to optimize and modify our model to visualize its layers through de-convolution. This paper analyzed the evolution of model features through layers and the part information learned by the higher layers. Based on these algorithms, we participate huawei Mo VAC2014competition and achieved the Best Performance Award.