Abstract
Current visual assessment of coating damage is performed by experienced inspector. This leads to variation in procedures and decisions among inspectors. Disagreement caused during assessment introduces issue of low reliability and repeatability of assessment results. A software that automatically identifies and determines percentage of defective areas in a given image is presented. Tested on a dataset of structural defects, outcome of the software is validated against answers provided by engineers and inspectors. The software requires an image of coating surface and hints on regions of interest as inputs. Users provide hints by clicking few times on regions they are interested (known as foreground) and not interested in (known as background). With built-in intelligence, the software scans the image and identifies defective areas by comparing colour difference between each pixel and hints user provided. If colour value of a pixel is found to be closer to hints in foreground, it is labelled as defective area and vice versa. Next, percentage of defective area in the image is determined and reported. Dataset of structural defects (http://xiscobonnin.github.io/resources/#datasets) with 73 images was used for experimental evaluation. These images come from wide range of engineering scenarios and contain different levels of coating damage. Wide variation that exists in the images would allow comprehensive understanding of performance of the software. Without prior learning, given each image, users provided hints by clicking few times (less than 10) on foreground and background. Machine (or computer) took over the analysis subsequently using algorithm developed based on colour information. Outcome from the software was validated against decisions made by engineers and inspectors. Results are promising with more than 90% of the defective areas successfully identified and in agreement with inspectors’ judgements. Reported percentage of defective areas for each image is also aligned with manually estimated value. This demonstrates that the developed tool is capable of improving process of visual inspection of coating surface. With current state of the art, user's hints have been proven to be very useful in guiding machine or algorithm to segment the image into foreground (region with coating damage) and background (irrelevant region). However, challenges remain for difficult situations such as colour of background being close to foreground and machine is unable to differentiate. Developed software is able to assess condition of coated surface quantitatively based on given image. This minimises the chance of argument over qualitative assessments during inspection. Having standardised tool for assessment would also allow comparison among existing cases. This serves as first step towards capture of experience in the industry through automation. Refinements such as threshold adjustment can be made subsequently to minimise difference between decision made by the software and experienced personnel.