A Domain-Level Transformation Scenario Classification Algorithm Based on Multilevel Feature Representation

Traditional scenarioscenario classification techniques mainly use SIFT features combined with BoW and SPM models for recognition, but this method is limited in accuracy and difficult to cope with variable scenario contents. To address this problem, this study proposes a new image scenario classification algorithm using multilevel feature representation. The algorithm constructs low-level and mid-level descriptions containing local semantics by sampling image blocks through a sliding window, extracting SIFT and CNN features, encoding them using the VLAD technique, and concatenating the features in order to form image descriptions. These descriptions are fused with the high-level semantics of the fully-connected layer to achieve the integration of local and global information and improve the accuracy of the image representation. Experiments on two commonly used datasets show that this fusion of multilevel features for image representation significantly improves the classification performance.