Proper visual identification of construction materials is critical for automated quality control in sustainable soil-based three-dimensional (3D) printing. Misidentification may compromise both structural performance and environmental goals. Non-contact material verification methods are needed to complement the rapid growth of sustainable extrusion-based construction. The purpose of this study is to build a vision-based framework to distinguish soil-printed elements from other similar-looking cement-printed and nonsoil-printed materials.
A data set of 3D-printed construction elements was created and expanded to six variants: Red, Green and Blue (RGB) and Grayscale, besides their processed and augmented versions. Multiple models were trained using a single ResNet50 backbone architecture, each for a different data set variant. An 80/20 train-validation split across 40 epochs was used in backbone transfer learning. Model performance was measured by accuracy, confusion matrices, class-wise precision, recall and F1-score.
Retaining RGB color significantly improved classification, achieving a validation accuracy of 86.12%. Grayscale variants remained competitive due to well-preserved textures, while offline data augmentation negatively affected performance in both color spaces. Error analysis showed that gray-toned soil and similarly colored cement samples continued to confuse each other.
This paper presents an early systematic analysis of preprocessing methods for vision-based material classification in soil-based 3D-printed construction. It develops a replicable computer vision workflow and suggests best preprocessing practices relevant to automated inspection and quality control in sustainable construction.
