Surface defect classification and detection on extruded aluminum profiles using convolutional neural networks

For decades, aluminum extrusion has been successfully applied in the manufacturing of profiles for the applications ranging from locomotives to skyscrapers. In recent years however, increasing profile complexity and the need for rapid production have lead to greater challenges for manufactures seeking rapid and robust production procedures. As a consequence, the occurrence of defects in extruded profile surfaces continues to create difficulties often requiring disposal of entire components. Hence, quality inspection of the profiles must be performed prior to packing in order to identify and appropriately manage defect-containing extrusions. Up until now, quality control in extrusion factories is primarily performed by the human eye due to its high performance in discriminating defect varieties. But human performance is cost intensive and furthermore prone to failure, especially when applied in high-throughput environments. On that account this paper proposes an approach in surface defect classification and detection, whereby a simple camera records the extruded profiles during production and a neural network architecture distinguishes between immaculate surfaces and surfaces containing a variety of common defects (surface defect classification). Furthermore, a neural network is employed to point out the defects in the video frames (surface defect detection). In this work, we show that methods from artificial intelligence are highly compatible with industrial applications such as quality control even under common industry constraints such as very limited data set sizes for training a neural network. Data augmentation as well as transfer learning are the key ingredients for training networks that meet the high requirements of modern production facilities in detecting surface defects, particularly when access to training sets is limited. Accuracies of 0.98 in the classification and mean average precisions of 0.47 in the detection setting are achieved whilst training on a data set containing as little as 813 images. Real-time classification and detection codes are implemented, and the networks perform reliably despite changes in lighting conditions and camera orientation.