基于改进Cascade R-CNN的两阶段销钉缺陷检测模型

无人机在输电线路巡检过程中会拍摄大量图片,自动识别无人机拍摄图片中存在的部件缺陷是无人机巡检的重要环节.其中销钉的缺陷由于目标较小且需要依赖上下文信息才能正确判断,识别难度较大.针对上述问题,提出了一种两阶段的销钉缺陷检测模型.首先使用Faster R-CNN (regin convolutional neural networks)模型提取出原始图像中的连接部位,再对提取出的每个连接部位进行缺陷识别.缺陷识别模型使用改进的Cascade R-CNN,该模型使用层级残差卷积模块代替骨干网络中的3×3卷积并使用路径聚合特征金字塔(PAFPN)代替原始网络中的特征金字塔结构,能够有效提取图片中的多尺度特征和上下文信息.最后将级联检测器的最后一级替换为double-head检测器,减少模型误报.实验结果表明,模型对销钉缺失及销钉脱出两类缺陷的平均识别精度能够达到81.2％,与原始的Cascade R-CNN相比提升了7.8％.

Two-stage pin defect detection model based on improved Cascade R-CNN

: UAVs will take a large number of pictures during the inspection of transmission lines. Automatic identification of component defects in pictures taken by UAVs is an important part of UAV inspections. Pin-level defects are difficult to identify due to the small target and the need to rely on context information for correct judgment. To solve the above problems, a two-stage pin defect detection model is proposed. First, the Faster R-CNN model is used to extract the connection parts in the original image, and then the defect recognition is performed on each connection part extracted. The defect recognition model uses an improved Cascade R-CNN, which uses a hierarchical residual convolution module to replace the 3x3 convolution in the backbone network and uses the path aggregation feature pyramid (PAFPN) to replace the feature pyramid in the original network, which can effectively extract the image Multi-scale features and contextual information. Finally, the last stage of the cascade detector is replaced with a double head detector to reduce false positives of the model. The test results show that the average recognition accuracy of the model for two types of pin missing and pin falling defects can reach 81.2% , which is 7.8% higher than the original Cascade R-CNN.