基于AW-Net的轮毂射线图像分割算法

Segmentation algorithm of wheel ray image based on AW-Net

针对传统方法下的汽车轮毂内部缺陷检测效率低、精度达不到工业标准的问题,本文提出了一种基于改进U-Net神经网络的轮毂X射线图像缺陷分割方法AW-Net。该方法通过三级跳跃连接的方式级联两个U型网络对图像特征进行深度提取。同时在跳跃连接的过程中融合注意力机制以解决小目标的变化情况容易被漏检的问题,并通过实验验证结合使用多种激活函数来实现更精准的轮毂X射线图像语义分割,增加网络的拟合能力,提高网络的鲁棒性。实验结果表明:改进后的算法在本文构建数据集的汽车轮毂内部缺陷的误判率为2.73%,漏判率为0,识别率达到93%以上,其分割精度高于传统图像分割网络全卷积网络(fully convolutional network, FCN)和U-Net,且本方法边缘分割更加平坦,满足现代轮毂内部缺陷无损检测的需要。

Aiming at the problems of low efficiency and accuracy in detecting internal defects of automobile wheels under traditional methods, and the accuracy is not up to industry standards, this paper proposes a method for segmentation of image defects in X-ray images of wheels based on improved U-Net neural network, AW-Net.This method cascades two U-shaped networks to extract image features in a three-level jump connection mode;at the same time, the attention mechanism is integrated in the jump connection process to solve the problem that the change of small targets is easy to be missed, and passes Experiments verify that a combination of multiple activation functions is used to achieve more accurate semantic segmentation of X-ray images of the hub, increase the fitting ability of the network, and improve the robustness of the network.The experimental results show that the improved algorithm has a false detection rate of 2.73%, a leakage rate of 0 and a recognition rate of more than 93% for the internal defects of automotive wheels in the data set constructed in this paper, and its segmentation accuracy is higher than that of traditional image segmentation networks, such as fully convolutional network(FCN) and U-Net, and the edge segmentation of this method is flatter and meets the needs of nondestructive detection of internal defects of modern wheels.