A difficulty of the visual inspection for translated and rotated objects is the limitation of recognition rate. This thesis proposes two inspection methods which are invariant for translations and rotations of the objects to be inspected.
The first method employs the fast Fourier transformation for an inspected object to achieve translation invariance. The method measures the rotated angles of the object, which are used to rotate the FFT images of the rotated objects. Rotation of the FFT images in such a manner results in a rotation invariant feature of the object under inspection. The feature is used to train a neural network to inspect a defection of objects. A self-organizing map is used to classify between defected and pure objects. Experimental results show that the recognition rate of the objects in arbitrarily rotated angles is about 95 percents.
The second method employs a template matching scheme. The polar transformation with respect to the center of the object is performed to achieve translation invariance. Rotation invariance is obtained by calculating the cross-correlation between the image under inspection and good images stored. A good/bad classification is made by comparing the correlation value with the predefined threshold value. Experimental results with the method show that the recognition rate is about 90 percents.
Two methods proposed here employ the translation and rotation invariance properties in classification between good and bad images. The recognition rates of both methods are higher than the general machine vision system developed in the CORE Lab.