Page 29 - Read Online

Page 22 of 28 Zhang et al. J Mater Inf 2024;4:34 https://dx.doi.org/10.20517/jmi.2024.64

According to the automatic programming method, the complete programs for image splicing can be
generated on the workflow as follows: Initially, the images undergo preprocessing which entails noise
reduction, grayscaling and rescaling. Subsequently, the feature detector is established using the SIFT method
to accommodate the high-resolution characteristics inherent in welding images. The feature detection phase
extracts distinctive features from each image, resulting in the generation of descriptors. The
feature-matching algorithms include brute-force matching, nearest neighbor matching and K-dimensional
tree (KD-tree) matching. The KD-tree matching is utilized in the research to achieve an optimal balance
between model efficiency and accuracy. The feature-matching algorithm discerns corresponding points
between images based on the descriptors, thereby constructing a robust set of paired feature points.
Furthermore, the random sample consensus (RANSAC) is employed to compute the homography matrix,
which reflects the correspondence among feature points. Image fusion techniques including weighted
averaging, Poisson blending and Laplacian pyramid blending facilitate seamless transitions in overlapping
regions. The suitably sized blank canvas is prepared onto which the stitched images are rendered. Finally,
post-processing enhances the quality of the stitched images through the cropping of extraneous areas and
the adjustment of brightness to improve detail visibility. The image splicing programs applied in this
research are available in the Supplementary Materials. The technology abandons traditional manual
marking methods and adopts the unsupervised stitching approach, enabling precise stitching of weld joint
images through automated detection and matching of feature points . The technology exhibits remarkable
[98]
adaptability, ensuring consistent performance across diverse scenarios, lighting conditions and sensor
configurations. Meanwhile, by combining the clarity of high-magnification microscopy with the broad field
of view offered by low-magnification microscopy, image stitching technology facilitates the acquisition of
clear and complete microstructural images of weld joints.

The schematic illustration of the longitudinal cross-section of tube plate welding is shown in Figure 20A to
depict the micrograph position of the weld seam vividly. Through the image-stitching process, lucid and
expansive images of welding microstructures are obtained, facilitating the effortless observation of
microstructural characteristics and the precise identification of welding defects. The majority of weld joint
morphologies in the welding test samples exhibit a defect-free appearance, as depicted in Figure 20B,
indicating stable control of welding parameters and favorable welding conditions. As illustrated in
Figure 20C, the stitched image reveals porosity defects in the weld, which may be attributed to wet
electrodes, moisture, oil, rust on the weldment or excessive V and currents. Figure 21 showcases the
s
morphology of the welding sample with crack defects in grayscale mode, which are caused by stresses
generated during the cooling process. Notably, the center of the stitched image exhibits a black-striped area
due to the absence of overlap between images captured by the camera, indicating that not all areas of the
sample surface were captured.

CONCLUSIONS
In the present work, an in-depth investigation is conducted into the welding formation and microstructural
characteristics of titanium alloys. First of all, the overall framework and specific structural design of the
CNN-based defect detection model are elucidated. Image enhancement techniques are applied to augment
the weld pool image dataset, and transfer learning methods are adopted to enhance model training
effectiveness. The performance of ResNet34, MobileNetV2, and ConvNeXt on welding defect datasets is
evaluated. The classification accuracy of the pre-trained ConvNeXt model, utilizing transfer learning, attains
an impressive 99.52% following 500 training iterations. In contrast, the accuracies of MobileNetV2 and
ResNet34 stand at 85.94% and 93.41%, respectively. Additionally, through the visualization technique of
t-SNE, the operational dynamics of the deep learning model in defect detection tasks are thoroughly
examined. Utilizing a systematic layer-by-layer feature extraction process, the model recalibrates the