基于改进卷积神经网络的管焊缝X射线图像缺陷识别方法

针对卷积神经网络(CNN)应用于焊缝探伤图像识别时，目标区域占比小，局部信息冗余，激活函数小于零时出现硬饱和区导致模型对输入变化较敏感、网络参数难以训练的问题，采用超像素分割算法(SLIC)和改进的ELU激活函数构建CNN模型进行焊缝探伤图像缺陷识别. 首先，在CNN模型中使用ELU激活函数，在缓解梯度消失时对输入噪声产生更好的鲁棒性，同时，利用SLIC算法对图像像素进行像素块处理的特点，增大焊缝探伤图像中感兴趣区域的占比，降低局部冗余信息，提高模型在训练过程中的特征提取能力. 通过对焊缝探伤图像感兴趣区域提取并与所述CNN模型进行对比试验. 结果表明，该方法在焊缝探伤图像特征提取、训练耗时及识别准确率方面较传统卷积神经网络有更好的表现.

X-ray image defect recognition method for pipe weld based on improved convolutional neural network

When convolution neural network (CNN) is applied to weld flaw detection image recognition, the target area is small, the local information is redundant, and the hard saturation region of activation function is less than zero, which makes the model sensitive to input change and difficult to train the network parameters. The super pixel segmentation algorithm (SLIC) and the improved ELU activation function are used to construct CNN model for weld flaw detection image defect recognition. First, the ELU activation function is used in the CNN model to generate better robustness to the input noise when the response gradient disappears, At the same time, the SLIC algorithm is used to deal with the pixels of the image, which increases the proportion of the region of interest in the weld flaw detection image, reduces the local redundant information, and improves the feature extraction ability of the model in the training process. Through the extraction of the region of interest of weld flaw detection image and the establishment of the CNN model described in this paper, the results show that the proposed method has better performance than the traditional convolution neural network in feature extraction, training time and recognition accuracy of weld flaw detection image.