High-dimensional multimedia classification using deep CNN and extended residual units

摘要

Processing multimedia data has emerged as a key area for the application of machine learning methods Building a robust classification model to use in high dimensional space requires the combination of a deep feature extractor and a powerful classifier. We present a new classification pipeline to facilitate multimedia data analysis based on convolutional neural network and the modified residual network which can integrate with the other feedforward network style in an endwise training fashion. The proposed residual network is producing attention-aware features. We proposed a unified deep CNN model to achieve promising performance in classifying high dimensional multimedia data by getting the advantages of the residual network. In every residual module, up-down and vice-versa feedforward structure is implemented to unfold the feedforward and backward process into a unique process. The hybrid proposed model evaluated on four datasets and have been shown promising results which outperform the previous best results. Last but not the least, the proposed model achieves detection speeds that are much faster than other approaches.