Fig. 4

Feature Pyramid Network (FPN) Implementation for Water Level Detection in Cementitious Samples. (a) The FPN integrates an EfficientNet-B2 backbone for hierarchical feature extraction, using skip connections to preserve spatial details. The decoder refines up-sampled feature maps through convolutional layers, generating high-resolution segmentation masks. (b) Input images, probability masks, and overlays illustrate accurate segmentation of water absorption regions. The sigmoid-activated “predict_step” method ensures precise binary mask predictions under varying conditions. (c) The impact of real and synthetic image datasets on segmentation accuracy (DoM) shows stability improvements with synthetic data and practical training with limited real data. Outliers appear as points below the main distribution. The data, adapted from Kabir et al.^20,21, is presented in a modified format.