Recognition of the condition of construction materials using small datasets and handcrafted features

We propose using handcrafted features extracted from small datasets to classify the conditions of the construction materials. We hypothesize that features such as the color, roughness, and reflectance of a material surface can be used to identify details of the material. To test the hypothesis, we have developed a pre-trained model to classify material conditions based on reflectance, roughness and color features extracted from image data collected in a controlled (lab) environment. The knowledge learned in the pre-trained model is finally transferred to classify material conditions from a construction site (i.e., an uncontrolled environment). To demonstrate the proposed method, 80 data points were produced from the images collected under a controlled environment and used to develop a pre-trained model. The pre-trained model was re-trained to adapt to the real construction environment using 33 new data points generated through a separate process using images collected from a construction site. The pre-trained model achieved 93%; after retraining the model with the data from the actual site, the accuracy had a small decrease as expected, but still was promising with an 83% accuracy.