Slim-neck by GSConv: a lightweight-design for real-time detector architectures

IF 2.9 4区计算机科学 Q2 COMPUTER SCIENCE, ARTIFICIAL INTELLIGENCE

Journal of Real-Time Image Processing Pub Date : 2024-03-29 DOI:10.1007/s11554-024-01436-6

Hulin Li, Jun Li, Hanbing Wei, Zheng Liu, Zhenfei Zhan, Qiliang Ren

引用次数: 0

Abstract

Real-time object detection is significant for industrial and research fields. On edge devices, a giant model is difficult to achieve the real-time detecting requirement, and a lightweight model built from a large number of the depth-wise separable convolutional could not achieve the sufficient accuracy. We introduce a new lightweight convolutional technique, GSConv, to lighten the model but maintain the accuracy. The GSConv accomplishes an excellent trade-off between the accuracy and speed. Furthermore, we provide a design suggestion based on the GSConv, slim-neck (SNs), to achieve a higher computational cost-effectiveness of the real-time detectors. The effectiveness of the SNs was robustly demonstrated in over twenty sets comparative experiments. In particular, the real-time detectors of ameliorated by the SNs obtain the state-of-the-art (70.9% AP₅₀ for the SODA10M at a speed of ~ 100 FPS on a Tesla T4) compared with the baselines. Code is available at https://github.com/alanli1997/slim-neck-by-gsconv.

Abstract Image

查看原文本刊更多论文

GSConv 的 Slim-neck：用于实时检测器架构的轻量级设计

实时物体检测对工业和研究领域意义重大。在边缘设备上，巨大的模型难以达到实时检测的要求，而由大量深度可分离卷积建立的轻量级模型也无法达到足够的精度。我们引入了一种新的轻量级卷积技术--GSConv，在保持精度的同时减轻了模型的重量。GSConv 在精度和速度之间实现了很好的权衡。此外，我们还在 GSConv 的基础上提出了一种设计建议--细颈（SNs），以提高实时探测器的计算成本效益。在二十多组对比实验中，SNs 的有效性得到了有力证明。特别是，与基线相比，SNs 改进的实时探测器达到了最先进水平（SODA10M 在特斯拉 T4 上以约 100 FPS 的速度运行时，AP50 为 70.9%）。代码见 https://github.com/alanli1997/slim-neck-by-gsconv。

本文章由计算机程序翻译，如有差异，请以英文原文为准。

求助全文

约1分钟内获得全文求助全文

来源期刊

Journal of Real-Time Image Processing COMPUTER SCIENCE, ARTIFICIAL INTELLIGENCE-ENGINEERING, ELECTRICAL & ELECTRONIC

CiteScore

6.80

自引率

6.70%

发文量

审稿时长

6 months

期刊介绍： Due to rapid advancements in integrated circuit technology, the rich theoretical results that have been developed by the image and video processing research community are now being increasingly applied in practical systems to solve real-world image and video processing problems. Such systems involve constraints placed not only on their size, cost, and power consumption, but also on the timeliness of the image data processed. Examples of such systems are mobile phones, digital still/video/cell-phone cameras, portable media players, personal digital assistants, high-definition television, video surveillance systems, industrial visual inspection systems, medical imaging devices, vision-guided autonomous robots, spectral imaging systems, and many other real-time embedded systems. In these real-time systems, strict timing requirements demand that results are available within a certain interval of time as imposed by the application. It is often the case that an image processing algorithm is developed and proven theoretically sound, presumably with a specific application in mind, but its practical applications and the detailed steps, methodology, and trade-off analysis required to achieve its real-time performance are not fully explored, leaving these critical and usually non-trivial issues for those wishing to employ the algorithm in a real-time system. The Journal of Real-Time Image Processing is intended to bridge the gap between the theory and practice of image processing, serving the greater community of researchers, practicing engineers, and industrial professionals who deal with designing, implementing or utilizing image processing systems which must satisfy real-time design constraints.