A Hybrid Approach for Moving Scene Analysis in Videos Using Discrete Wavelet Transform and Kernelized Region-Based CNNs

Abstract

An innovative framework is designed with the combination of Discrete Wavelet Transform (DWT) and Kernelized Region Based Convolutional Neural Networks (KR-CNN) for extracting and classifying the moving video frame recognition. In this, the DWT is mainly utilized for decomposing the video frames in various representations which captures both the spatial domain and frequency domain features that are required for dynamic understanding of the scenes. Then, the proposed KR CNN is used to extract the fine grain region based features and also to classify them which shows that the KRCNN is important for the network to develop the complex spatial relationships. This proposed combination is experimentally analyzed with the datasets for achieving the higher accuracy of classification of 96.8% with better parametric values. To justify the novelty of the proposed method, this combination is also tested and compared with conventional CNN methods and RCNN methods and the results are clearly depicting that the DWT and KRCNN combination proves themselves by providing greater improvement in classification accuracy with reduced computation time which provides to make use of the proposed method in the real time video analysis which is significant requirement in the filed of multimedia based applications.