Joints Steering Gear Detection Based on Semantic Segmentation

Aiming at the problem of inaccurate detection of joints steering gears in the industrialized inspection of automobiles, a method for joints steering gear detection based on image semantic segmentation is proposed. In this method, considering that this learning tasks are usually easy to overfit, appropriate regularization is needed to generalize well. In data augmentation, we randomly cover the input square area and randomly fill the corresponding square areas of other data into the area which is covered. Adopting encoder-decoder structure to obtain semantic segmentation graph, in downsampling we use dilated convolution to increase feature map size. And we also use label smoothing and adjust the loss function to improve the ability to predict the test set more accurately. Experiments show that this method is effective and has a good generalization for the detection problem of joints steering gear grooves.