PSI - Issue 66

Andrii Kompanets et al. / Procedia Structural Integrity 66 (2024) 388–395 Author name / Structural Integrity Procedia 00 (2025) 000–000

391

4

3. Method We make use of an encoder-decoder network based on convolutional neural networks (CNN) to segment images of cracks in steel bridges. The encoder-decoder network maps each pixel of the input image to a value of 0 (background) or 1 (crack). It can be noted that throughout the paper, segmentation maps are visualized with inverted colors - white background and a black crack. To enhance the performance of the neural network, we make use of the recent ConvNext (Liu, 2022) architecture as the network encoder. Furthermore, we enhance the architecture further by incorporating spatial and channel squeeze-and-excitation (scSE) layers (Roy, 2018). Lastly, we modify several of the final (high scale) layers by introducing skip connections to improve the handling of fine-grained details in the images. The overall encoder-decoder network architecture is illustrated in Figure 2.

Fig. 2: Scheme of the proposed encoder-decoder network.

3.1. Experiments In this study, the ADAMW algorithm (Loshchilov, 2017) is used to optimize the neural network weights. The initial learning rate is set to λ e=1 = 0.001 with an exponential decay schedule, defined as: λ = λ e=1 γ e (1) where λ is the learning rate, γ =0.99, and e is the current epoch number. To make better use of the ConvNext pre trained weights, we implement a stage-wise learning rate decay technique: λ stage = λ k N+1-n (2)