提出了一种针对处理散乱堆叠物体的改进方法.在YOLOv5模型中采用了加权双向特征金字塔网络(BiFPN)替代路径聚合网络(PANet),结合Gfocal损失函数,使得漏检和误检问题得到有效改善,平均精度均值(mAP)mAP@0.5达到了90.1%.利用 Mask R-CNN进行目标物体分割,使用轻量化的 Mobilenetv3替代ResNet101主干网络以减少参数量,同时借用CFNet思想加强特征融合机制,使得分割精度提高至92.1%.通过级联改进后的YOLOv5和改进后的 Mask R-CNN,算法在实时性和精确性上得到了平衡,在有效感兴趣区域(region of interest,ROI)中提取准确的物体形状信息.与单独使用实例分割算法相比,检测速度提升了1 s.实验证明所提出的算法不仅提高了推理速度,还提高了分割精度,解决了复杂堆叠场景下物体特征提取效果差且检测速度慢的问题.
Segmentation algorithm for scattered stacked objects based on cascade network
This paper proposes an improved method for handling scattered stacked objects.In the YOLOv5 model,the BiFPN feature pyramid is used to replace PANet,and combined with the Gfocal loss function,the problem of missed detection and false detection is effectively improved,and mAP@0.5 reaches 90.1%.Mask R-CNN is used for target object segmentation,the lightweight Mobilenetv3 is used to replace the ResNet101 backbone network to reduce the number of parameters,and the CFNet idea is used to strengthen the feature fusion mechanism,increasing the segmentation accuracy to 92.1%.By cascading the improved YOLOv5 and the improved Mask R-CNN,the algorithm achieves a balance between real-time performance and accuracy,and extracts accurate object shape information in the effective region of interest(ROI)area.Compared with using the instance segmentation algorithm alone,the detection speed is increased by 1 s.Experiments have shown that the algorithm proposed in this article not only improves the inference speed,but also improves the segmentation accuracy,and solves the problem of poor object feature extraction and slow detection speed in complex stacking scenes.
NETL
NSTL
万方数据
