Error Assessment of Machine Vision Techniques for Object Detection and Evaluation

As image-based assessment practices become increasingly present in the field of civil engineering, so do the interpretation errors. In machine vision, these errors are typically represented by quantitative metrics stating the presence of error. In order to provide a comprehensive assessment of the state of a structure and structural elements, errors should be interpreted and rationally integrated into decision making. In this paper, a performance-based approach to include the source and impact of the errors in machine vision object detection techniques is presented. Structural element evaluation is used as an example. The approach involves an adaptation of the error-domain model falsification method developed previously. More specifically, an automated method in reinforced concrete column damage state estimation based only on visually observed damage characteristics is used as a study case. Predicted damage state(s) (eleven in total, DS0-DS10) are compared with the measured damage state of the column surface image that has been retrieved by way of the automated machine-vision method. Ultimately, a set of damage states is provided as equally possible solutions for the input image of the column surface. This analysis helps focus efforts to reduce errors on image characteristics that are the greatest source of ambiguity. In this study, lower errors associated with determining damage states 0 to 5 would be the most helpful.