基于改进像素差分神经网络的碳管多层编织成形视觉检测

Visual detection of carbon tube multilayer braiding process based on improved pixel difference neural network

针对碳管预制体编织成形过程中编织网轮廓缺陷人工难以及时发现的问题,提出了多层碳管预制体编织网编织过程中的轮廓视觉检测方法。在原始像素差分神经网络的基础上进行改进,采用SimAM模块实现了网络的轻量化和更好的注意力关注效果,使用平衡交叉熵损失解决了碳管编织网数据样本分布不均匀的情况,同时,提出一种逐像素的编织网轮廓比对方法来判断编织网状态,以此判断编织网是否出现缺陷。试验结果表明:与原始像素差分神经网络相比,改进的像素差分神经网络对编织网轮廓提取结果的ODS和OIS分别提高了2.4个百分点和2.6个百分点。认为:改进的像素差分神经网络对于碳管编织网轮廓实现了更为精确的提取,更加适用于碳管多层编织成形视觉检测任务;逐像素比对方法对编织网轮廓缺陷可以实现准确及时的判断。

Aiming at the problem that it was difficult to find the defects of the braided mesh in time in the process of braided carbon tube prefabrication.The visual detection method of carbon tube multilayer braiding process was proposed.Based on the original pixel difference neural network,SimAM module was used to realize the lightweight of the network and better attention effect,the balanced cross-entropy loss was used to solve the problem of uneven distribution of carbon tube braided mesh data,and a pixel-by-pixel braided mesh outline comparison method was proposed to judge the braiding net state.The experimental results showed that compared with the original pixel difference neural network,the improved pixel difference neural network improved the ODS and OIS by 2.4 percentage points and 2.6 percentage points respectively.It is considered that the improved pixel difference neural network can extract the outline of the braided mesh more accurately,which is more suitable for the visual detection task of the braided mesh.The method of pixel-by-pixel comparison can accurately and timely judge the outline defects of braided mesh.