Detection of scratches on car rims with classical computer vision approaches does not produce good results, because the rims are of different shapes, made of different materials and in different colors. Scratches are also in different shapes, colors and sizes. Scratches are often poorly visible and there is also dirt on rims, which further impedes visibility of them. This requires the use of more powerful tools, like convolutional neural networks that have been experiencing rapid development over the last few years. In thesis we analyze possibility of scratch detection on car rims with deep neural network. Segmentation of each point in the input image is required for detection purposes. Because the scratches are small compared to the entire region of the rim, we decided to use a fully convolutional network U-Net. For a purpose of the thesis, a collection of annotated pictures was prepared, which can be a starting point for further research.
The developed model successfully detects scratches, despite the small learning set. On an unseen test set prepared for evaluation purposes only, it achieved 62.8\,\% accuracy using the mIoU method. With further improvements and refinements, our scratch detector would also be suitable for industrial use.
|
