Key points are not available for this paper at this time.
Crack identification in concrete structures plays a crucial role in maintaining structural integrity and preventing progressive deterioration. However, most existing image-based approaches are limited to surface-level analysis and struggle to provide reliable depth estimation. Their performance degrades under noise, illumination imbalance, and overlapping pixels, and they also lack a balanced feature-extraction mechanism capable of distinguishing true crack features from background variations. To address these limitations, this article introduces a Crack Depth Detection Method (CDDM) using a weighted Neural Network (wNN). The method extracts textural features in terms of intensity and saturation and constructs a column-matrix representation to isolate variations in pixel arrangement and distribution. The wNN recurrently analyses overlapping pixels and assigns adaptive weights to improve depth estimation accuracy and reduce misclassification under varying image conditions. The proposed CDDM achieves an overall detection accuracy of 93.038%, improving accuracy and precision by 15.54% and 13.48%, respectively, while reducing misdetection by 7.01% compared to existing methods. Instances where the model reaches 100% accuracy occur only within small, controlled test subsets and do not imply universal or absolute accuracy; they simply reflect perfect classification within that specific subset. The overall performance remains consistently at 93.038% across full-scale testing
R. et al. (Sun,) studied this question.