Research On Automatic Detection Algorithm Of Wheel Hub Defect Based On Convolutional Neural Network

Defect detection on the inside of the wheel hub is very important for the driving safety of the car.The use of fully automated wheel inspection systems can effectively improve the efficiency of the factory’s detection of internal wheel defects and optimize the production process of automotive wheels.Defect detection algorithm is the core of fully automated wheel defect detection system.The existing wheel hub defect detection algorithms have shortcomings such as slow speed,missed judgment and misjudgment.Aiming at the above problems,this paper proposes a fast detection algorithm for wheel hub defects based on convolutional neural network.At the end of this paper,the effectiveness of the algorithm is verified by the dataset.This paper firstly explores the feasibility of using deep learning technology to detect the internal defects of the wheel hub.Next,this paper starts from the classic image semantic segmentation model U-Net and studies the influence of deepening the network depth on the experimental results.The new model introduces a gated attention mechanism based on the deepened U-Net model and the model is fine-tuned and improved using various activation functions and convolution kernels of different sizes.We call this modified model AW-Net.This improved model we propose can achieve an m Io U index of 0.899 on the wheel defect dataset in this paper,which solves the problem that small defects inside the wheel are easily missed.It is found that the AW-Net model improves the detection accuracy in the wheel defect segmentation experiments,but there are problems of model redundancy and misclassification of noise.Meanwhile,in order to realize a faster wheel defect detection algorithm,this paper follows the end-to-end idea of AW-Net model and proposes a lightweight defect semantic segmentation model named Effi-Deep Lab.The new model is proposed after analyzing the characteristics of convolutional methods such as conventional convolution,atrous convolution,and depth-wise convolution.After comparing the popular lightweight feature extraction networks,Efficient Net is selected as the feature extraction network.We redesigned the ASPP structure after combining the actual characteristics of the internal defects of the wheel.The model is further improved by multi-scale fusion at the decoding end to improve the model accuracy.In order to solve the problem of misclassifying noise as defects that occurs in AW-Net experiments,we not only use conventional data enhancement techniques,but also add noise to the images and use image enhancement techniques to achieve better data enhancement.The experiments show that the above approach can deepen the learning of the defective region by the model and improve the robustness and detection effect of the model.We established a wheel image dataset containing multiple types of defects.We established a wheel image dataset containing multiple types of defects and implemented semantic segmentation of three types of defects: shrinkage,shrinkage and cracking.The comparison experimental results show that the algorithm proposed in this paper can identify defects quickly and accurately.