Scene Recognition of Surveillance Data using Deep Features and Supervised Classifiers

Precise labeling of an image based on its semantic description is quite challenging task and has its significant applications in surveillance area. Majority of scene classification techniques during past few decades have targeted low level feature by handcraft engineering or unsupervised feature extraction techniques. In this paper, we aim to categorize scene classes for surveillance systems by exploiting deep convolutional features to manifold projection along with supervised classification algorithms. A topology is constructed to depict high dimensions of convolution heat-maps to 128D salient features. Parameters of pre-trained network are tuned to precisely fit with the output of our problem. Experimental results depict that our methodology is more robust and competitive as compared to state of the art methods.