Dynamic texture using deep learning
Rishabh Bansal, Arun Singh Pundir, Balasubramanian Raman · 2017
Identifying object in a dynamic scene is one of the main problems in computer vision. This is directly related to solving recognition problem for dynamic texture. Recognizing dynamic texture has become a fundamental problem to understand natural video content. It is a powerful technique for recognizing natural scenes such as fire, waves and smoke. Methods which exist today suffer from various problems such as camera motion, varying illumination etc. To solve these problems, we propose a deep learning based method, which have shown considerable success in image classification. One of the major requirements for using deep-learning techniques is a large dataset which is not available for dynamic texture. Hence, in this paper, we have given our own dataset divided over 11 categories with each category consisting of atleast 50 videos. We have used a pre-trained CNN model which is fine tuned on our dataset to extract features and use it as an input to SVM/Random Forest classifiers. This way transferring the learning from a pre-trained CNN reduces the computation cost. Results section have shown the robustness and effectiveness of given method on our own dataset.