Surface defect detection of aluminum alloy wheels is an important part of the production logistics process of aluminum alloy wheel enterprises. Through on-site investigation, it is found that in the traditional solution, the surface defect detection of aluminum alloy wheels is to locate and mark the defects through the naked eyes of workers at the fixed stations of the production line. However, due to the complex structure of automobile aluminum alloy wheels, the manual detection method is inefficient and the workload is large, it is increasingly difficult for this detection method to meet the higher and higher requirements of rapid, accurate and stable aluminum alloy wheel production process. In order to realize the automation and intelligence of production logistics equipment, improve the efficiency of the production logistics process of aluminum alloy wheel manufacturing enterprises, and reduce labor costs, in response to the above situation, combined with deep learning technology, and field research on aluminum alloy wheel defect detection tasks, a deep Learned online detection algorithm for defects in aluminum alloy wheels.
main tasks as follows:
1. Process design of defect detection system for aluminum alloy wheel production line. This system includes a complete online defect detection process for aluminum alloy wheels, including image acquisition, image preprocessing, image defect detection of aluminum alloy wheels, and so on.
2. Establish a database of surface defects of aluminum alloy wheels. The specific process is to collect the surface defect images of the aluminum alloy wheels through industrial cameras on the production line, then clean the data, and mark the defect images through professional marking software, and finally get a defect image database with accurate markings.

3. Identify fuzzy images. The anti-blur algorithm used is implemented based on a generative countermeasure network based on deep learning, so the calculation speed is slow, and only a part of the collected images are blurred. In order to improve the efficiency of the defect detection algorithm system, before removing the image blur , Design an algorithm to identify fuzzy images, if the detection is fuzzy, it will be sent to deep learning to eliminate image fuzzy network. If it is recognized as a clear image, the defect detection of the aluminum alloy wheel image will be carried out directly.

4. Eliminate the blurring of the image on the surface of the aluminum alloy wheel. Before detecting the defect images of aluminum alloy wheels, the defect images need to be processed while the aluminum alloy wheels are being transported on the production line. Therefore, the collected images of aluminum alloy wheels will have motion blur. Aiming at this problem, this paper proposes a deep adversarial learning algorithm to eliminate motion blur without requiring accurate motion information.

5. Image defect detection of aluminum alloy wheels. Implemented the deep learning target detection algorithm and improved it based on the aluminum alloy wheel defect task. On the basis of the Faster-RCNN target detection algorithm, the SE module was added, and ROI-Pooling was replaced by ROI-Align, FPN multi-scale feature fusion network With three improvements, a model that can locate and classify surface defects of aluminum alloy wheels is finally obtained. The final realization of the aluminum alloy wheel image defect detection algorithm can detect two kinds of defects, defect points and scratches, and give the defect location and defect type at the same time. On the 300 surface defect image data sets of aluminum alloy wheels, a recall rate of 94% and a prediction accuracy rate of 88% can be achieved for all defect areas.