Community Research and Development Information Service - CORDIS

FP5

MONOTONE Report Summary

Project ID: G1RD-CT-2002-00783
Funded under: FP5-GROWTH
Country: United Kingdom

Algorithms for defect detection

Defect detection in stochastic textures via Parzen window (Mean Shift) method:
A non-parametric, unsupervised method. Colour pixel clustering performed to produce stack of binary images. Stack describes spatial colour distribution of given image. Parzen window test performed at each pixel to measure density and distribution regularity. Coherence map results indicating uniformity of texture in acquired image. Morphological and thresholding techniques applied to extract possible defects. Threshold established during training phase of algorithm. Good results indicating vibration lines, smudges and misprints. Method is difficult to extend to random textures with structural elements and computationally unsuitable for real-time application at this time.


Defect detection and localisation in pseudo-random textures via texem models:
Novel texture analysis model developed: texem model. Assuming each image is constructed from a set of sub-image patches of various sizes, possibly overlapping, then images of the same product contain similar base textural elements: texems. Given an initial training set of defect free images, a set of various sized image patches are extracted. Similar patches are grouped in a multidimensional feature space and clusters extracted via a Gaussian mixture model utilising an Expectation and Maximisation algorithm. Given a new image, extract a small patch at each pixel position and compare it to the set of learnt texems. A multi-scale approach is utilised to reduce computation costs. Candidate defective regions from different scales are combined to generate a final localised defect map. Method achieved good results but is computationally unsuitable for real-time application at this time.


Defect detection and localisation in pseudo-random textures via colour texem models:
Texem model extended to model colour images, initially via separate image eigenchannels. Given an initial set of defect free images, a reference eigenspace is derived such that decomposed images preserve major colour features. Grey level texem analysis is then applied independently to each decomposed eigenchannel. Defect candidates from each channel are combined to form final defect map. However, interaction between image channels ignored, e.g. vectorisation leads to causal effect that is not necessarily present within original image data. Therefore texem model extended to 3D. Improved performance noted, particularly when defects chromatic in nature. Method achieved good results but is computationally unsuitable for real-time application at this time.


Defect Detection via Frequency Space Analysis:
Generic defect detection technique for textured tiles via frequency space was investigated. A defect free image is split into patches and for each patch a Fourier transform calculated and a feature vector extracted. Statistical methods are applied to determine if a given patch differs significantly from its corresponding model patch. Initial results showed promise, however, method not applicable to pseudo-random textures and not viable for real-time implementation. Specific case of vibration lines then investigated. Training set of ideal tiles processed in frequency domain and best representative model profile extracted. Given new tile, profile extracted and compared to model via Chi squared similarity metric. Low order similarity statistics derived during model creation used to create threshold to determine if new profile significantly different. Initial results were promising and the algorithm runs in real-time. Colour mapping stage implemented to decrease suppression of subtle tonal variations. However, few data sets exist for testing and validation due to infrequent nature of defect.

Defect detection and localisation for pseudo-random tiles via affined templates (Jigsaw):
Method reconstructs texture space from multiple tile images, similar to solving a jigsaw puzzle. Correlation used to ‘stitch’ new pieces into texture space via hierarchical sub-sampled pyramid of images. Initial 2D position estimators via exhaustive search, histogram-based hash table indexing and Fourier-based phase correlation examined. For each level thereafter a 5D affined transform is derived with previous transform as basis for current search. Final model pyramid rendered via vector directional stack smoothing to provide resilience to noisy inputs. Newly acquired images located within model as in training, once located dynamic local window differencing, segmentation and a novelty metric derived. User-defined threshold applied to determine if region defective. Algorithm detected and localised defects and end users can tune algorithm sensitivity. Method performs in real-time and is equally applicable to fixed pattern tiles.

Related information

Reported by

University of Bristol
MVB, Woodland Road, Clifton
BS8 1UB BRISTOL
United Kingdom
See on map
Follow us on: RSS Facebook Twitter YouTube Managed by the EU Publications Office Top