改进 RetinaNet 的绝缘子精确定位研究

接触网图像中绝缘子部件的自动精确定位是绝缘子故障检测的基础,绝缘子在接触网图像中存在倾角,采用水平框进 行检测难以精确契合目标。 针对此问题,提出一种改进 RetinaNet 的绝缘子精确定位方法。 首先利用高效 Ghost 模块代替原特 征提取网络中的卷积操作获得多尺度特征图,减少模型计算量;其次将注意力机制嵌入网络中,抑制次要特征对目标检测的影 响;然后引入旋转框作为模型的预测框实现绝缘子精确定位,降低冗余背景噪声的干扰;最后重新定义训练过程中的正负样本, 解决了添加旋转框导致学习错误样本的问题。 实验结果表明,该方法可以精确地定位绝缘子,抑制冗余背景信息,与原算法相 比检测精度提高 2. 8%,检测速度为 25. 6 FPS,参数量减少 42. 8%,具有良好的检测性能。

Research on insulator accurate location based on improved RetinaNet

Automatic and accurate location of insulator components in catenary images is the basis of detecting insulator fault. The insulators in the catenary images are incline-angled, so it is difficult to locate the object accurately by using horizontal box. To solve this problem, an insulator accurate location approach was proposed based on improved RetinaNet. To begin with, the efficient Ghost module was adopted to replace the convolution operation in the original feature extraction network to obtain multi-scale feature maps and reduce the computational burden of the model. Next, in order to suppress the influence of secondary features on object detection, the attention mechanism was embedded in the network. Then, the rotating box was introduced as the prediction box of the model to realize the accurate location of insulators and reduce the interference of redundant background noise. Finally, the positive and negative samples were redefined in the training process. By doing so, the problem of learning wrong samples that caused by adding rotating box was resolved. Experimental results demonstrate that the proposed approach featuring good detection performance can locate the insulator accurately and prevent redundant background information. Compared with original algorithm, the detection accuracy increases by 2. 8%, the detection speed reaches to 25. 6 FPS, and the number of network parameters reduces by 42. 8%.