Scholarlink

<p>Integrating object detection into Flexible Printed Circuit(FPC) production promotes autonomous monitoring, bolstering safety, augmenting efficiency and quality, and cutting costs. However, object detection for the FPC production process is a complex multi-scene task with inter-object logical ties across scenes. To tackle this, this paper proposes a Cross-scene Object Relationship-drive Object Detection method(COR-OD), which extracts, models, and applies the production temporal sequence of objects within the same scene and the causality between objects across different scenes. The method employs a two-stage training approach. Multi-scene object images are sent to the base-detector to obtain visual features, which are then fed into the constructed object temporal causal cascade Mamba network(OTCC-Mamba) to derive relational feature transformation results. These results are then fused with visual features and output to the detection head for the final detection outcome. Additionally, we embed a relational attention module into the relational feature transformation process to enable the network to focus more accurately on correct relational expressions. The method’s performance, verified on the self-built MSO-FPC dataset, shows significant improvements. The AP50 rose by 4.0 to 72.7, and the mAP increased by 3.6 to 52.5, enabling the recognition of object relationships in multi-scene FPC production. Code available at <a href="https://github.com/FutureIAI/COR-OD">https://github.com/FutureIAI/COR-OD</a>.</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

Cross-scene object relationship-driven object detection for flexible printed circuit production

  • 期刊论文

  • 发表日期：

    2025年10月11日

查看全文
  • Jiu Dai,
  • Guangzhu Chen,
  • Xiaojuan Liao,
  • Linmao Xu

摘要

Integrating object detection into Flexible Printed Circuit(FPC) production promotes autonomous monitoring, bolstering safety, augmenting efficiency and quality, and cutting costs. However, object detection for the FPC production process is a complex multi-scene task with inter-object logical ties across scenes. To tackle this, this paper proposes a Cross-scene Object Relationship-drive Object Detection method(COR-OD), which extracts, models, and applies the production temporal sequence of objects within the same scene and the causality between objects across different scenes. The method employs a two-stage training approach. Multi-scene object images are sent to the base-detector to obtain visual features, which are then fed into the constructed object temporal causal cascade Mamba network(OTCC-Mamba) to derive relational feature transformation results. These results are then fused with visual features and output to the detection head for the final detection outcome. Additionally, we embed a relational attention module into the relational feature transformation process to enable the network to focus more accurately on correct relational expressions. The method’s performance, verified on the self-built MSO-FPC dataset, shows significant improvements. The AP50 rose by 4.0 to 72.7, and the mAP increased by 3.6 to 52.5, enabling the recognition of object relationships in multi-scene FPC production. Code available at https://github.com/FutureIAI/COR-OD.