Hybrid Neural Network Architecture with Adaptive Loss Function for Semantic Segmentation of Residential Building Defects
Yevgeniy M. Mamatov and
Yaroslav Yu. Golovko
Additional contact information
Yevgeniy M. Mamatov: Phazotron Corporation - Research Institute of Radio Engineering
Yaroslav Yu. Golovko: Belgorod State National Research University
Economics. Information Technologies, 2026, vol. 53, issue 2, 416-423
Abstract:
The article is focused on the problem of automating crack detection and segmentation in residential buildings using deep neural network architectures. A hybrid model named HybridUNet is proposed, combining an EfficientNet-B0 encoder pre-trained on ImageNet, a transformer bottleneck for modeling long-range spatial dependencies, and spatial attention modules in the decoder to suppress background noise. An adaptive loss function is applied, dynamically adjusting the contribution of Dice, Focal, and Boundary components depending on the epoch: the Dice loss weight is fixed at 1, the Focal loss weight linearly decreases from 1 to 0.5, while the Boundary loss weight increases from 0.1 to 1. This allows the model to first focus on the global defect structure and then refine boundaries. Comparison with baseline U-Net models (with ResNet34 and EfficientNet-B0 encoders) and DeepLabV3+ is performed on an open crack dataset. Experiments show that the proposed architecture achieves the best crack IoU (0.6029) and F1 score (0.7369) on the test set, outperforming DeepLabV3+ by 6.6% in IoU. The adaptive loss improves boundary accuracy, while the transformer block ensures robustness to noise and partial occlusions. The obtained results confirm the applicability of hybrid architectures with adaptive loss tuning for automated inspection of residential building stock
Keywords: semantic segmentation; cracks; HybridUNet; vision transformer; adaptive loss function; EfficientNet; spatial attention (search for similar items in EconPapers)
Date: 2026
References: Add references at CitEc
Citations:
There are no downloads for this item, see the EconPapers FAQ for hints about obtaining it.
Related works:
This item may be available elsewhere in EconPapers: Search for items with the same title.
Export reference: BibTeX
RIS (EndNote, ProCite, RefMan)
HTML/Text
Persistent link: https://EconPapers.repec.org/RePEc:eul:ecoinf:v:53:y:2026:i:2:p:416-423
DOI: 10.52575/2687-0932-2026-53-2-416-423
Access Statistics for this article
Economics. Information Technologies is currently edited by Elena A. Stryabkova
More articles in Economics. Information Technologies from Belgorod State University
Bibliographic data for series maintained by Tatyana M. Khomenko ().