Visual recognition and classification of videos using deep convolutional neural networks

Abstract

Classification of videos based on its content is one of the challenging and significant research problems. In this paper, a simple and efficient model is proposed for classification of sports videos using deep learned convolution neural networks. In the proposed research, the gray scale variants of image frames are employed for classification process through convolution technique at varied levels of abstraction by adapting it through a sequence of hidden layers. The image frames considered for classification are obtained after the duplicate frame elimination and each frame is further rescaled to dimension 120x240. The sports videos categories used for experimentation include badminton, football, cricket and tennis which are downloaded from various sources of google and YouTube. The classification in the proposed method is performed with Deep Convolution Neural Networks (DCNN) with around 20 filters each of size 5x5 with around stride length of2 and its outcomes are compared with Local Binary Patterns (LBP), Bag of Words Features (BWF) technique. The SURF features are extracted from the BWF technique and further 80% of strongest feature points are employed for clustering the image frames using K-Means clustering technique with an average accuracy achieved of about 87% in classification. The LBF technique had produced an average accuracy of 73% in differentiating one image frame to other whereas the DCNN had shown a promising outcome with accuracy of about 91% in case of 40% training and 60% test datasets, 99% accuracy in case of 60% training an 40% test datasets. The results depict that the proposed method outperforms the image processing-based techniques LBP and BWF.