基于改进Mask R-CNN模型的电力场景目标检测方法

Object Detection Method of Electric Power Site Based on Improved Mask R-CNN

为了解决电力施工现场中安全帽佩戴情况以及危险区域行人入侵检测问题,提出一种基于改进Mask R-CNN模型的目标检测方法。首先依据迁移学习策略对Mask R-CNN主干网络进行参数初始化,以提取图像基本特征;然后引入特征金字塔结构进行自下而上的特征图提取,完成多尺度特征融合;接着,通过多尺度变换方法对区域推荐网络进行调整,获取锚点进行回归计算完成检测实验;最终对结果进行分析评价,多目标平均准确率达到了95.22%。将改进后的Mask R-CNN模型用于监控视频分析,针对监控视频像素过低问题,加入拉普拉斯算法锐化边缘,精准率提高到90.9%,验证了拉普拉斯算法对低质量监控视频检测的有效性。

In order to solve the problems of helmet wearing in power construction site and personnel intrusion detection in dangerous areas,an object detection method based on improved Mask R-CNN model was proposed.Firstly,the parameters of Mask R-CNN backbone network were initialized to extract image basic features using transfer learning.Then the feature pyramid network was used to extract the bottom-up feature map and complete the multi-scale feature fusion.Then,the regional recommendation network was adjusted by multi-scale transformation method,and anchors were obtained to complete the detection experiment.Results show that average accuracy of multi-objective is 95.22%.The improved Mask R-CNN model was applied to the analysis of surveillance video.Aiming at the problem of low pixel of surveillance video,laplacian sharpening was added to improve the video quality.The experimental results show that the accuracy rate is improved to 90.9%,which verifies the effectiveness of laplace algorithm for low quality surveillance video detection.