Video Concept Detection Using SVM and CNN

  Jeevan J.Deshmukh,  Nita S.Patil,  Dr.Sudhir D.Sawarkar

In fast growing digital world, with very high speed internet videos are uploaded on web. It becomes need of system to access videos expeditiously and accurately. Concept detection achieves this task accurately and is used in many applications like multimedia annotation, video summarization, annotation, video indexing and retrieval. The execution of the approach lean on the choice of the low-level visible features employed to show the key-frames of a shot and the preference of method used for extracting the feature. The syntactic differences among low-level features abstracted from video and human analysis of the video data are linked by Concept Detection System. In this proposed work, a set of low-level visible features are of greatly smaller size and also proposes effective union of Support Vector Machine(SVM) and Convolutional Neural Networks (CNNs) to improve concept detection, where the existing CNN toolkits can abstract frame level static descriptors. To deal with the dataset imbalance problem, dataset is partitioned into segments and this approach is extended by making a fusion of CNN and SVM to further improve concept detection. To increase efficiency and to get the result within lesser time the existing systems lags and so using the video reader to extract the frames. Frame undergoes Hu_moments and HSV histogram to produce feature vector for classification. This paper makes two contributions first, the two classifiers SVM and CNN are separately trained on data set and this enriches efficient result. The accuracy of each classifier is individually calculated. Second, the fusion of two classifiers is performed to efficiently detect the concepts in test dataset. After the fusion of two classifiers, the accuracy is calculated. The proposed framework using fusion of SVM and CNN gives effective video concept detection. Accuracy is used as measure to evaluate the system performance for UCF 101dataset. The fusion of CNN and SVM classifiers provides better results in comparison with individual classifier. The proposed framework is validated on standard UCF 101 dataset using accuracy as predictive measure.

Keywords- Support vector machine; Video Concept Detection; Convolutional Neural Network; Key Frame Extraction; Feature Extraction

Unique Identification Number - IJEDR1903009
Page Number(s) - 47-52
Pubished in - Volume 7 | Issue 3 | July 2019
DOI (Digital Object Identifier) -    http://doi.one/10.1729/Journal.22720
Publisher - IJEDR (ISSN - 2321-9939)

  Jeevan J.Deshmukh,  Nita S.Patil,  Dr.Sudhir D.Sawarkar,   "Video Concept Detection Using SVM and CNN", International Journal of Engineering Development and Research (IJEDR), ISSN:2321-9939, Volume.7, Issue 3, pp.47-52, July 2019, Available at :http://www.ijedr.org/papers/IJEDR1903009.pdf