Detection of Scratches on Cars by Means of CNN and R-CNN

Abstract

Failure detection systems have become important not only in production processes, but nowadays there is a need for their implementation in various daily areas, for example, for the detection of physical damages that a car may present in a parking lot, in order to provide to the client with the assurance that their personal property will not be affected inside the lot. This paper presents an algorithm based on convolutional neural networks and a variant of these using regions (CNN and R-CNN), which allows to detect scratches in a car. In the first instance, the capture of one of the sides of a conventional car is done, an R-CNN is designed to extract only the region of the image where the car is located. After this procedure, the extracted region is divided into multiple sections, and each of the sections is evaluated in a CNN to detect in which parts of the vehicle the scratches are located. With the test images, a precision percentage of 98.3% is obtained in the R-CNN, and 96.89% in the CNN, demonstrating in this way the robustness of the Deep Learning techniques implemented in the detection of car scratches. The processing times of each one of the algorithm stages corresponding to the R-CNN and the classification of the sections in the CNN were 1.6563 and 1.264 seconds respectively.