The Quality Estimation from Objects Group Image of Non-uniform Texture using Wavelet Hierarchy

Abstract

We describe classification into overall quality of objects produced in groups, corresponding with the surface appearance of the objects in non-uniform texture images. A method applied with a hierarchy by the Wavelet transform and a neural network (NN) is proposed. This method does not require mechanical picking or arranging before taking images, because the method directly uses an image of overlaping objects in the producing.
Reliable regions for classification are discriminated from unreliable regions using spatial localization of the Wavelet hierarchy through the reconstruction. Useful components for classification are extracted from the regions using frequency decomposition of the hierarchy. Wavelet channels (frequency bands) for the extraction are optimized in order that a brightness histgram from the extracted components can reflect difficulty according to classes. NN learns classification using the histgram based on judgment of a human expert, and yield an overall quality class of the whole image.
Experimental results prove practicality of the proposed method in the classification performance and in the processing speed.