保存先:
| 第一著者: | |
|---|---|
| フォーマット: | Recurso digital |
| 言語: | 英語 |
| 出版事項: |
Zenodo
2023
|
| 主題: | |
| オンライン･アクセス: | https://doi.org/10.5281/zenodo.10403761 |
| タグ: |
タグ追加
タグなし, このレコードへの初めてのタグを付けませんか!
|
目次:
- <p>As deep neural networks continue to drive advancements in computer vision tasks, researchers are increasingly focusing on developing powerful deep neural network models to handle large volumes of data. As the size of DNN models grows, their inference process becomes computationally expensive, which limits the real-time applications of DNNs. In response to this challenge, we propose a universal framework called "Locality Sensing and Optimized Neural Network for Video Processing SPEED-UP." This framework utilizes locality sensing to accelerate the video query processing, reducing the computational costs of DNNs in video evaluation and achieving a threefold reduction in inference time.</p> <p>This framework can automatically sense the similarity between two input frames within a given input video. It enables us to process input videos using a specialized processing method that incurs significantly lower computational costs compared to traditional DNN inference, which involves frame-by-frame detection. Within the highlighted temporal locality information across frames, the Yolov5 algorithm can achieve a two to threefold speedup. Furthermore, we employ neural network compression techniques to reduce the size of the neural network without sacrificing minimal accuracy, ultimately compressing the neural network volume by half. The simultaneous use of these two methods validates the effectiveness of our research.</p>