电梯钢丝绳视觉缺陷识别方法研究

Research on visual defect identification method of elevator wire rope

根据电梯钢丝绳检测存在人工检测效率低下、精度不佳的问题,基于树莓派4B设计了一套钢丝绳视觉缺陷识别装置,同时搭载改进后的YOLOv7算法:将主干部分卷积层替换成蛇形动态卷积;对卷积层进行残差连接,使得网络可以直接学习残差;将网络结构中的Upsample替换为CARAFE,可以更准确地恢复细节和纹理信息;最后在网络的SPPCSPS模块中添加Biformer注意力机制,以改进对图像中不同区域的重要性建模能力,提高模型对图像细节和关键目标的识别能力,以及加强对全局的考虑。经实验,改进后的算法模型应用于电梯钢丝绳视觉缺陷检测,平均精度m AP@0.5达到了95%,相比改进前算法提高了3.1%,满足了钢丝绳的检测要求。

According to the low efficiency and poor accuracy of manual detection in elevator wire rope detection,a set of visual defect identification device for wire rope is designed based on Raspberry Pi 4B,and the improved YOLOv7 algorithm is also equipped.The convolution layer of the trunk part is replaced by dynamic snake convolution,and the convolution layer is connected with residuals,so that the network can learn residuals directly.Replacing Upsample with CARAFE in the network structure can restore details and texture information more accurately.Finally,the Biformer attention mechanism is added to the SPPCSPS module of the network to improve the ability to model the importance of different regions in the image,improve the ability of the model to identify image details and key targets,and strengthen the consideration of the overall situation.The improved algorithm model is applied to the visual defect detection of elevator wire rope,and the average accuracy m AP@0.5 reaches 95%,which is 3.1 percentage points higher than that before the improvement,and meets the requirements of related wire rope detection.