基于改进YOLACT的堆叠零件实例分割算法

Instance Segmentation of Cluttered Mechanical Parts Based on Improved YOLACT

为了解决堆叠环境下零件实例分割精度差的问题,提出了一种改进YOLACT算法。通过在主干网络中C3和C4层引入多级特征融合与通道注意力机制模块(MLCA),优化了特征提取的精度。为了在保证图像同时获取多感受野信息,采用上下文特征金字塔模块(AC-FPN)结构替代传统FPN金字塔,获取更多感受野,以准确完成预测。通过自制堆叠零件数据集完成网络训练与实验。对比实验表明,改进后的YOLACT算法在未明显提升运行时间的基础上,相较原算法表现出更优的检测与分割效果。

In order to address the issue of poor instance segmentation accuracy for parts instances in a cluttered environment,an improved YOLACT algorithm is proposed.Multi-level feature fusion and channel attention mechanism modules(MLCA)are introduced into the C3 and C4 layers of the backbone network,optimizing the precision of feature extraction.At the same time,to ensure the image acquires multi-scale field-of-view information,an advanced contextual feature pyramid module(AC-FPN)structure replaces the conventional FPN pyramid to capture broader receptive fields,facilitating accurate prediction.The network was trained and tested using a custom-made dataset of stacked parts.Comparative experiments demonstrate that the improved YOLACT algorithm,without significantly increasing the run time,exhibits superior detection and segmentation performance compared to the original algorithm.