Intelligent Brick Detection: A One-Class Deep Learning Framework for Construction Material Identification

Accurate recognition of construction materials is crucial for quality assurance, safety, and monitoring in the building industry. Conventional computer vision approaches typically rely on supervised classification, which requires both positive and negative training samples. However, collecting representative non-brick data is challenging due to the diversity of materials and environmental conditions. To overcome this limitation, this study proposes a one-class learning framework for brick detection, designed to operate using only positive brick images.A dataset of 1,080 brick images was sourced from Kaggle’s “Bricks under Construction or Old Building Houses” repository and divided into training (70%), validation (20%), and testing (10%) sets. Feature embeddings were extracted using MobileNetV2, a pre-trained convolutional neural network, producing compact and discriminative image representations. These embeddings were then used to train a One-Class Support Vector Machine (SVM) with an RBF kernel to model the feature distribution of bricks and identify outliers.Experimental results demonstrate the effectiveness of the proposed method. At the chosen threshold (5th percentile of training scores), the model achieved 93.06% recall on the validation set and 92.59% recall on the test set, ensuring reliable acceptance of brick images. Robustness was further confirmed through 5-fold cross-validation, yielding an average recall of 87.17% ± 3.08%. Decision score histograms and Recall vs Threshold analysis indicated that recall consistently exceeded 90% across a broad range, highlighting system stability.The study establishes that one-class learning provides a practical and scalable solution for construction material identification, eliminating the need for comprehensive negative datasets and offering strong potential for future multi-material and real-time applications.