管道漏磁信号的自动检测是无损检测领域一直未解决的难题，随着计算机视觉技术的发展与应用，利用深度学习实现漏磁信号的自动检测成为研究热点。本项目重点解决基于多尺度注意力机制的管道漏信号自动检测关键问题，主要内容包括：(1) 针对管道漏磁信号成像问题，建立变周期采样数据伪彩色成像模型，设计基于拉普拉斯与多尺度结构元素的图像增强模型，强化目标边缘特征；(2) 针对漏磁信号分类问题，以卷积神经网络为基础框架，加入稀疏自编码和图像熵相似度约束规则对卷积核优化训练，提高网络特征提取能力和区分性表达能力；(3) 针对漏磁信号自动识别问题，设计以SSD网络为基础的注意力模型，通过加入空洞卷积和注意力残差模块，改进现有模型中对小缺陷识别能力和定位能力差的问题，通过加入3d实时在线检测模块，完善系统的实时性。本项目的研究有助于解决管道漏磁信号的高精度自动识别问题。因此，具有至关重要的理论意义和工程价值。

The automatic detection of pipeline magnetic leakage signal has remained the unresolved problem in the field of the non-destructive inspection. With the development and application of computer vision technology, it is one of the research hotspots to achieve the automatic detection of magnetic leakage signal by deep learning. This project aims to solve the main issues associated with pipeline leakage signal automatic detection using artificial intelligence. This includes: (1) The pseudo-color image transformation model of the variable period sample data was built with reference to the imaging of pipeline magnetic leakage signals, while the image enhancement model based on multiscale structural elements and Laplace were designed to further improve image resolution; (2) As for the classification of magnetic leakage signals, added the pre-training model of sparse own coding and similarity constraint rules of image entropy to perform optimization training with convolutional neural network as the basic framework, and conducted the optimization training for convolutional kernels, thus enhance the edge features of the target; (3) The deep learning model, based on an SSD network, was designed in conformance to the automatic identification of magnetic leakage signals. The dilated convolution and attention residual module were added to enhance the capacity of defecting the minor defects in existing modules, thus increasing the positioning accuracy of the targets to be detected. The real-time 3D online detection module was added to perfect the real-time performance of the system. The research study for this project plays a significant role in dealing with problems with low precision and efficiency of magnetic leakage defect identification. Therefore, it is crucial both in theoretical significance and engineering value.