Attentive Convolution Network-Based Video Summarization

The availability of smart phones with embedded video capturing mechanisms along with gigantic storage facilities has led to generation of a plethora of videos. This deluge of videos grasped the attention of the computer vision research community to deal with the problem of efficient browsing, indexing, and retrieving the intended video. Video summarization has come up as a solution to aforementioned issues where a short summary video is generated containing important information from the original video. This paper proposes a supervised attentive convolution network for summarization (ACN-SUM) framework for binary labeling of video frames. ACN-SUM is based on encoder–decoder architecture where the encoder is an attention-aware convolution network module, while the decoder comprises the deconvolution network module. In ACN-SUM, the self-attention module captures the long-range temporal dependencies among frames and concatenation of convolution network and attention module feature map result in more informative encoded frame descriptors. These encoded features are passed to the deconvolution module to generate frames labeling for keyframe selection. Experimental results demonstrate the efficiency of the proposed model against state-of-the-art methods. The performance of the proposed network has been evaluated on two benchmark datasets.