A Flexible Machine Vision System for Small Parts Inspection Based on a Hybrid SVM/ANN Approach

Loading...
Thumbnail Image
Date
Authors
Joshi, Keyur
Keyword
Machine Vision , Inspection , Hybrid , Flexible System , Classification , Pattern Recognition , Quality Experiments , Support Vector Machine , Artificial Neural Network
Abstract
The automated inspection and sorting of parts is a common application of Machine Vision (MV). The sorting of parts is possible only after reliable classification. The goal of this thesis was to develop and validate a flexible MV system that can reliably classify small parts. Support Vector Machines (SVMs) and Artificial Neural Networks (ANNs) are popular choices as classification algorithms. Classifiers developed from supervised algorithms perform well when trained for a specific application with known classes. Their drawback is that they are considered inflexible as they cannot be easily applied to a different application without extensive retuning. Moreover, for a given application, they do not perform properly if there are unknown classes. Classifiers developed from semi-unsupervised algorithms can work with unknown classes but cannot work with multiple known classes. A novel solution to these limitations has been developed using a hybrid two-layered approach with supervised SVMs, semi-unsupervised SVMs and supervised ANNs. With the hybrid approach as the basis for the classifier, a flexible MV system was designed with four key characteristics: 1) cost effective hardware, 2) a realistic and manageable image database, 3) an effective image conditioning process and 4) a comprehensive features library. Four hybrid classification methods were developed and tested: 1) semi-unsupervised SVM followed by supervised SVM (USVM-SSVM), 2) supervised SVM followed by semi-unsupervised SVM (SSVM-USVM), 3) semi-unsupervised SVM followed by supervised ANN (USVM-SANN) and 4) supervised ANN followed by semi-unsupervised SVM (SANN-USVM). The target performance criteria for the system was an accuracy of 95% with 0% false positives. To validate the system and to demonstrate its flexibility, experiments were conducted with two hardware setups, three applications (gears, connectors, coins) and five sets of high quality images with known/unknown classes. The effect of image quality was studied by digitally blurring and dimming the conditioned images. It was found that SANN-USVM gave the best results and exceeded the target performance criteria. A software package known as FlexMVS for Flexible Machine Vision System was written to evaluate the hybrid approach and to enable easy execution of the image conditioning, feature extraction and classification steps.
External DOI