An Improved U-Net Image Segmentation Network for Crankshaft Surface Defect Detection

Crankshaft is one of the mechanical components of the vehicle engine, and quality control of it holds significant importance in the production line. In this paper, a vision-based system was developed to detect apparent structural defects on the crankshaft surface. By examining the different approaches in computer vision tasks, the semantic segmentation technique was chosen to solve this problem. In the first stage, a dataset consisting of 400 crankshaft experimental images with structural defects such as scratch, pitting, and grinding were collected. Then, the Convolutional Neural Network (CNN) with MobileNet architecture was trained to detect apparent defects, and an Intersection over Union (IoU) evaluation criteria of 64.7% was obtained. In the third stage, some image processing techniques were used to increase the performance. By applying the DexiNed edge detection filter on the train-set images, the IoU was increased by 8.4%. Considering the importance of this issue in the automotive industry, it has been tried again to boost the performance by augmenting the dataset images. On the other hand, this can also prevent overfitting of the model. By training the model under the same conditions as the previous stages, the IoU in this stage increased by 13.2% and reached 86.3%.