A Machine Vision Based Inspection Approach for Casting Defects Detection

Abstract

Imaging based inspection schemes have lately gone quite widespread for faults isolation and quality regulatory framework. The proposed work introduces a novel approach for the detection of casting defects that abscond conventional inspection in foundries. This research incorporates various image processing techniques that are compared to identify the most suitable feature extractor and feedforward back propagation method. The techniques used include image morphology for structural element based image processing, Grey Level Co-Occurrence Matrix (GLCM) for analysis of the castings’ textures by considering the spatial relationship of image pixels, Laws Texture Energy Measures (LTEM) that applies convolution masks to calculate the texture energies of defects and lastly implementation of Tamura textures for fault detection as these features significantly imitate human visual perceptions. The classification of the defects has been done using feed forward back propagation neural network. The proposed algorithm takes features from the input imagery including entropy, mean, contrast, correlation, energy and homogeneity thus making considerable fault segregation. Similarly, energy values obtained from each of the LTEM kernel are analysed in the neural network for results, however, the attempted feature extraction in this case of textural unevenness of the castings has been best examined using the Tamura textural features. Tamura’s capability to idealise the human visual sense with respect to coarseness, contrast and directionality achieves the desired results when their energy maps are passed through a feedforward back propagation neural network. Overall average classification accuracy of 86.3% is achieved when Tamura based features are applied as input to the artificial neural network and best validation performance for the training data is found to be 0.2787. In comparison, the GLCM and LTEM based features achieved best validation performances at 0.3260 and 0.3544, respectively. Total average classification accuracy achieved from their respective testing dataset is 83.5% and 83.1%, respectively. Therefore, the proposed technique demonstrates promising results by using Tamura textural features to characterize casting defects using a feed forward Back Propagation Neural Network (BPNN).