Classification of shaped tubes and profiles for laser cutting machines

New technologies provide high-performance and high accuracy on computer-based systems. Image processing and machine learning are developed to solve complex problems. Image processing methods are highly effective in cases where manpower is insufficient or slow in mass production, quality control, automation area. Image processing is capable to increase productivity and efficiency. In cases where visual control is not possible and the ability to detect with sensors is impossible, the use of image processing methods may serve low cost, well-optimized and safe processes. In this thesis, a solution that is classifying shaped tubes-profiles that are used in the metal fabricating industry is presented. Fully automatic pipe-profile loading systems are used in laser cutting machines. Automatic detection of the loaded tube and profile in these systems is not possible with traditional methods. Due to this inadequacy, the automatic loading systems in the existing machines allow only one type of tube-profile to be loaded at the same time. This problem leads to a waste of time. On the other hand, the inability to detect tubes-profiles automatically requires the use of many complex structures in mechanical design. In this thesis, the classification process of special-shaped tube-profiles is carried out with the application developed by taking into account the industrial conditions. It is aimed to perform a wide range of classification processes by creating a database for thirty different tube-profile shapes. The data in the created databases were reproduced via image processing methods. It is aimed to increase the accuracy rate in deep network training by applying the developed image preprocessing algorithm to the images in the database. Training and tests were carried out for different database combinations using the created databases and GoogleNet and ResNet deep networks. The GoogleNet model, which is trained with a database of processed and unprocessed images, is successful, providing higher accuracy with a shorter training time compared to the deep network training results and the tests made in the real-time application. The model accuracy average was measured as 99.73%, and the classification prediction time was observed as 89 milliseconds on average. The model predicted the correct class in all of the predictions made for 30 classes in real-time tests. Models trained with ResNet are not preferred due to the high training time, the high error rate in classification predictions, and approximately 2 times slower classification time.