17 August 2022 in IEEE Transactions on Instrumentation and Measurement
A method based on the multilayer feature fusion multiscale GhostNet (MFMSGN) is proposed to improve the accuracy of magnetic flux leakage (MFL) image recognition of pipeline corrosion defects. First, a multiscale ghost module is constructed to make it more suitable for practical applications, which can obtain a large number of multiscale features through a small amount of calculation. Then, a parallel structure is used instead of the traditional method of improving accuracy by deepening the model depth. In this article, we introduce an adaptive spatial feature fusion (ASFF) method to fuse features of different resolutions to widen the width of the network and apply it to downstream classification tasks. Finally, a lightweight and efficient model is proposed. The results of the experiment showed that the accuracy of the eight-layer MFMSGN for MFL corrosion defect identification is better than ResNet50 and slightly inferior to ResNet101, but the computational effort is much smaller.
29 May 2023 in IEEE Transactions on Industrial Informatics
To solve the problem of the difficult and slow detection of small defects in magnetic flux leakage (MFL), we propose a fast MFL small defect detection network (FSDDNet). First, we introduce COMSOL multiphysics (COMSOL) data augmentation method that utilizes COMSOL simulation software to obtain high-resolution images of defects, which allows the network to capture complete defect features. Furthermore, this method addresses the issue of the relatively monotonic nature of the MFL defect dataset used in the experiment. Deep-learning networks usually use stride = 2 or max pooling to downsample the feature map, but this method will make the feature map lose some information, and small targets will lose more fine-grained information. Therefore, we introduce an SPD-Conv method to downsample the feature map, which can effectively avoid the loss of information. Meanwhile, an improved C3 network is introduced in the backbone network of FSDDNet. It greatly decreases the computational effort of the network and improves the detection speed. Finally, we add a small target detection head, which effectively improves the small target accuracy. FSDDNet is improved on the basis of YOLOv5, and after the above improvements, FSDDNet obtains very good results in the problem of MFL small defect detection. Experiments show that the accuracy of this algorithm is 95.2% when IOU = 0.5 and the latency is 7.9 ms.
23 August 2023 in IEEE Transactions on Instrumentation and Measurement
To apply deep learning algorithms to magnetic flux leakage (MFL) detection, we propose an anchor-free pipeline MFL image detection method (AFMFLDM) that can simultaneously combine low latency and high accuracy. The algorithm is modified based on CenterNet. The anchor-free target detection algorithm does not need to design the anchor box size compared to the one-stage and two-stage target detection algorithms, and there is no nonmaximum suppression (NMS) process, which reduces the computational effort. Then, the backbone of this algorithm is selected as a modified PP-LCNet, which replaces the normal convolution with a depthwise separable convolution. It is supplemented with a technique of adjusting parameters to form a network similar to MobileNetV1, which ensures low computational effort and high accuracy compared with the popular feature extraction networks. Finally, a feature fusion module based on receptive field convolution (FFRF) is introduced to improve the detection accuracy. The experimental results show that the accuracy of the algorithm is 95.6% when the intersection over union (IOU) is greater than 0.5, and the inference time is 8.7 ms, which can meet the actual demand of pipeline MFL detection.