| 陈丹,令陈佩,刘瑞瑜.新型DRNet结合EIoU的遮挡目标分割模型[J].电子测量与仪器学报,2025,39(8):209-217 |
| 新型DRNet结合EIoU的遮挡目标分割模型 |
| Novel DRNet occlusion target segmentation model combined with EIoU |
| |
| DOI： |
| 中文关键词: 遮挡物体 实例分割 DRNet Cluster EIoU NMS EIoU损失 |
| 英文关键词:occluded objects instance segmentation DRNet CEIoU NMS EIoU loss |
| 基金项目:国家自然科学基金(61973085)、福建省自然科学基金(2022J01114)项目资助 |
|
|
|
|
| 摘要点击次数: 613 |
| 全文下载次数: 682 |
| 中文摘要: |
| 实例分割是计算机视觉领域的重要研究方向，但由于遮挡问题的存在，使得该任务仍然没有得到充分探索。针对目前算法对遮挡物体的分割检测效果不佳，容易出现误检漏检问题，在Mask R-CNN框架基础上，提出一种新型双向残差网络（DRNet）结合EIoU的遮挡目标分割模型。首先，提出一种DRNet代替原有ResNet网络，使用更少的BN层和ReLU层取代传统Conv-BN-ReLU结构，利用传统卷积和深度可分离卷积串行连接增强图像感受野特征，通过跳跃连接减轻网络随深度增加出现退化问题，提升网络表征能力；其次，使用CEIoU NMS算法代替原有NMS算法，通过聚类思想有效处理重叠边界框抑制问题，引入EIoU评估指标增加边界框几何信息，更加精准地描述边界框之间的相似程度，减少网络对遮挡物体边界框的错误抑制；最后，使用EIoU损失替换原有Smooth L1损失，加速网络收敛速度，提升边界框检测精度。在公共COCO 2017数据集上进行预训练，再在不同程度的遮挡数据集上进行实验。实验结果表明，相比较于原网络，所提分割算法在COCO 2017数据集上Box AP和Mask AP分别提升了1.7% 和1.3%；在遮挡数据集上对遮挡物体边界框检测精度和掩码分割精度均有明显提升，证实该方法对遮挡物体分割的有效性。 |
| 英文摘要: |
| Instance segmentation is an important research direction in the field of computer vision, but the existence of the occlusion problem still prevents this task from being fully explored. To address the poor segmentation detection of occluded objects by current algorithms, which are prone to the problems of misdetection and omission, a novel duplex residual network （DRNet） is proposed, combining the EIoU occluded target segmentation model with the Mask R-CNN framework. First, DRNet is proposed to replace the original ResNet network, using fewer BN and ReLU layers to replace the traditional Conv-BN-ReLU structure, utilizing the conventional convolution and depth-separable convolution serial connection to enhance the image sensory field features, and mitigating the degradation problem of the network with the increase of the depth by the hopping connection. Second, the CEIoU NMS algorithm is used instead of the original NMS algorithm to effectively deal with the overlapping bounding box suppression problem through the clustering idea, and the introduction of the EIoU evaluation index increases the bounding box geometric information, which more accurately describes the degree of similarity between the bounding boxes, and reduces the network’s erroneous suppression of the bounding boxes of the occluded objects. Finally, the EIoU loss is used to replace the original Smooth L1 loss to accelerate the network convergence speed and improve the bounding box detection accuracy. In this paper, we first conduct pre-training on the public COCO 2017 dataset and experiments on different degrees of occlusion datasets, and the results show that compared with the original network, the proposed segmentation algorithm improves the Box AP and Mask AP by 1.7% and 1.3% on the COCO 2017 dataset, respectively; and both the bounding-box detection accuracy of the occluded object and the mask segmentation accuracy on the occlusion dataset are significantly improved on the occlusion dataset, confirming the effectiveness of the method for occluded object segmentation. |
| 查看全文 查看/发表评论 下载PDF阅读器 |
|
|
|
