On approximate nearest neighbor field algorithms in template matching for surface quality inspection
A. Quesada-Arencibia et al.
|Titel||On approximate nearest neighbor field algorithms in template matching for surface quality inspection|
|Buchtitel||EUROCAST 2013 Computer Aided Systems Theory Extended Abstracts|
|Verlag||IUCTC Universidad Las Palmas|
Surface quality inspection is applied in the process of manufacturing products for which the surface appearance is crucial for the product quality and customer acceptance. Typical examples are woven fabrics like industrial textiles. The predominating approaches used in this ?eld are feature-based. This means that these methods rely on the choice of a subset of computed features from a virtually in?nite pool of possible features and combinations of them. Further on these features provide the base for decision algorithms like support vector machines. The appropriateness of the feature selection and the con?guration of the selected decision algorithm substantially depend on the characteristics of the surface texture and possible defects. The state-of-the-art of this approach is represented by Gabor-?lter banks and grey-level co-occurrence matrices (GLCM). These algorithms require careful tuning, which might cause corresponding con?guration and maintenance e?ort. Additionally, more complex and sophisticated features can only be realized at demanding computational costs. Recently  investigates an alternative approach that utilizes template matching in the context of regular or near-regular textured surface inspection. The approach follows the basic idea: ”If I can ?nd a good registration for a test patch in a reference image, then this patch is defect free”. This approach is not feature-based. There are only a limited number of con?guration parameters left like a sensitivity parameter. All other inherent parameters for example the window size are determined by statistical analysis. However, the performance and the e?ectiveness of using template matching in this context relies on the choice of the similarity measure. The approach proposes to utilize the so-called discrepancy measure as ?tness function for template matching and shows its feasibility, but leaves open questions particularly concerning the localization accuracy. The method proposed in this paper aims at re?ning the mentioned approach in terms of defect localization and is motivated from a di?erent ?eld of image processing, namely structural image editing. For structural image editing so-called nearest neighbor ?eld (NNF) algorithms are an active ?eld of research. NNFs provide relations between patches in a source image and similar patches in a reference image. There are e?ective approximate NNF algorithms like Patchmatch which are successfully applied for in-painting and image reconstruction applications. In this paper we look at NNF from the registration point of view based on the discrepancy measure as ?tness function. Particular properties of the discrepancy measure, mainly monotonicity and a Lipschitz property make this measure suitable for registration and the combination with NNF. The proposed paper evaluates whether approximate NNF algorithms in general are suitable as acceleration technique for template matching especially in the context of regular or near-regular textured surfaces. A comparison will be done with the original Patchmatch, the Coherence Sensitive Hashing and a new Patchmatch variant using the discrepancy measure. All experiments will be carried out on standardized test samples. To demonstrate the relevance in industrial applications the surface inspection algorithm of will be enhanced with an approximate NNF computation to improve defect localization.