Thumbnail Image

Image processing and machine learning for segmentation and classification of aluminium extrusions

Inventory management is one of the most important components of administration in a large-scale manufacturing business. Architectural Profile Limited produces tonnes of aluminium profiles every single day and have over one thousand different extrusions available. Many of the profiles look almost identical to each other and are very difficult for even a highly experienced employee to distinguish. Sending out an incorrect extrusion will cost time and money for both the manufacturer and the customer. Automation using machine vision is already a prominent tool used in factory environments worldwide. The ability to automate repetitive tasks frees up time for higher skill tasks and reduces manual data entry errors. A real time inventory monitoring system would allow for quick and confident decision making and reduce order problems due to out of date stock numbers. The aim of this project is to explore whether classification of extrusion profiles from images is possible and what factors ensure an accurate classification system. To achieve the aim, new data was acquired that captures a range of variation such as different angles, positions and lighting. In this thesis two main paths were explored, segmentation followed by classification of binary images using generic Fourier descriptors, and direct classification from colour images. This thesis presents a qualitative and quantitative analysis of the different approaches, and discusses the challenges and limitations found with the data. The methods are compared and recommendations are highlighted regarding the next steps that aid creation of an automated aluminium extrusion inventory management system. Presented is a pipeline of extracting features using techniques such as Gabor filters, Gaussian blur, and edge detectors. The output filters are used to train a random forest pixel classifier to segment the extrusions. Perfect reference binary images generated via a semi-manual threshold process are used to train a Logistic Regression classifier, and the test images were automatically segmented and classified. This classifier makes predictions on the segmented images, and the testing accuracy achieved was 88%. Instances that were misclassified had poor segmentations, in most cases from uneven illumination on the extrusion end. The other promising method is training convolutional neural networks using transfer learning. VGG-19 was trained with only the dataset acquired and achieved a testing accuracy of 89%, while ResNet-50 was trained using data augmentation and achieved a testing accuracy of 90%. The resulting confusion matrices show misclassifications between extrusions that are very similar in shape and size. Further work will improve the performance of the classifiers by fine-tuning the parameters and carrying out parameter searching. Further investigation should be carried out given more requirements for the application, such as the processors and exact lighting conditions.
Type of thesis
The University of Waikato
All items in Research Commons are provided for private study and research purposes and are protected by copyright with all rights reserved unless otherwise indicated.