基于分割与深度学习网络的复杂电气图纸元件识别

Complex Electrical Drawing Component Recognition Based on Segmentation and Deep Learning Networks

摘要:

针对复杂电气图纸的像素稀疏性、尺寸差异性、数量差异性导致的元件漏检、误检问题，提出了一种基于分割与深度学习网络的元件识别方法。首先，构建图纸分割算法以减小图纸与元件的尺寸差异。其次，基于YOLOv5网络提出一种四尺度检测机制，增加两条特征传输路径，获取表征元件细节的极浅层特征图。同时，改进初始锚框选取方式，以重叠面积、距离、角度、宽高度4个因素表征定位损失，改善网络定位效果，提高网络收敛速度。在包含17种典型元件的数据集上验证了该方法的有效性，实验结果表明，该方法的平均均值精度可达96.7 %，比原始网络提高了21.5 %，网络训练速度也明显优于其他算法，具有较好的综合识别性能。

Abstract:

To address the issues of component omission and misdetection in complex electrical drawings， which are caused by the pixel sparsity， size discrepancies and variations in quantity， a component recognition method based on segmentation and deep learning network is proposed. First， a drawing segmentation algorithm is developed to reduce the size difference between drawings and components. Next， based on the YOLOv5 network， a four-scale detection mechanism is proposed， which incorporates two transmission paths for feature maps to capture shallow feature maps that characterize detailed component information. Additionally， the method or selecting the initial anchor frame is improved， and the localization loss is defined by four factors： overlap area， distance， angle and width-height ratio， which improves the network's localization accuracy and convergence speed. Finally， the effectiveness of the proposed method is validated on a dataset containing 17 typical components. The experimental results show that the average mean accuracy of the method can reach 96.7 %， a 21.5 % improvement over the original network. Moreover， the method shows significant improvements in training speed compared to other algorithms， resulting in a better overall recognition performance.