Scholarlink

<p>Visual motion processing is a fundamental function of the visual system, extensively studied across species. Despite advances in understanding neural mechanisms, motion perception remains incompletely understood. Guided by a correlation-based paradigm, we propose a bio-inspired feedforward network of spiking neurons. This network employs the Hodgkin-Huxley model and spatiotemporal receptive field correlations, mimicking the direction-selective responses observed in the visual cortex. The network generates a motion map that delineates the boundaries of moving objects based on neuron firing rates. We evaluate the performance of our network across diverse video sequences from established datasets, confirming the model’s capability for accurate moving boundary detection. Comparative evaluations against state-of-the-art spiking neural networks underscore its superior performance, notably achieving 89% accuracy on the CDnet-2014 dataset. Furthermore, we apply our model to motion segmentation, achieving state-of-the-art results on four benchmark datasets: Davis-16 (83.83%), Davis-2017 (86.1%), YouTube-VOS-18 (87.4%), and SegTrackv2 (83.30%). These results confirm the model’s robust capability to distinguish foreground objects from the background.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

A Correlation-Based Spiking Neural Network for Motion Boundary Detection and Segmentation in Dynamic Visual Sequences

  • 期刊论文

  • 发表日期：

    2025年11月19日

查看全文
  • Hayat Yedjour,
  • Abdelkader Haddag,
  • Dounia Yedjour

摘要

Visual motion processing is a fundamental function of the visual system, extensively studied across species. Despite advances in understanding neural mechanisms, motion perception remains incompletely understood. Guided by a correlation-based paradigm, we propose a bio-inspired feedforward network of spiking neurons. This network employs the Hodgkin-Huxley model and spatiotemporal receptive field correlations, mimicking the direction-selective responses observed in the visual cortex. The network generates a motion map that delineates the boundaries of moving objects based on neuron firing rates. We evaluate the performance of our network across diverse video sequences from established datasets, confirming the model’s capability for accurate moving boundary detection. Comparative evaluations against state-of-the-art spiking neural networks underscore its superior performance, notably achieving 89% accuracy on the CDnet-2014 dataset. Furthermore, we apply our model to motion segmentation, achieving state-of-the-art results on four benchmark datasets: Davis-16 (83.83%), Davis-2017 (86.1%), YouTube-VOS-18 (87.4%), and SegTrackv2 (83.30%). These results confirm the model’s robust capability to distinguish foreground objects from the background.