Fabric Defect Detection Based On Bidimensional Empirical Mode Decomposition

The advent of fabric defects impair fabric quality, appearance and property, consequently diminish the profit of manufacturers, so the detection of fabric defect is a vital procedure during the the textile corporations’ quality controlling work. However, the manual way of fabric inspection executed in the great majority of domestic textile companies is limited by human abilities in terms of vision and brain and leads to a low speed, poor detection rate, high objective and intensive labour expense. Additionally, the inspection results by experts cannot be utilized as input data for the manufacture administration. Therefore, the human way of defect detection has already been unsuitable for the modern textile manufacture, and replacing the human fabric defect detection with a machine-vision-based one is an inevitable trend of automation and informatization for the textile industry. While researches on automatic fabric inspection has been done over 30 years both abroad and interiorly, there was not a widely available type of fabric inspection machine, which was essentially caused by a lack of self-adaptivity in the existing detection algorithms on large numbers and various types of defects.This study aims at presenting a highly self-adaptive fabric defect detection algorithm based on machine vision. By analyzing the drawbacks of existing fabric defect detection algorithms, bidimensional empirical mode decomposition(BEMD) was introduced as the core of the proposed algorithm. The BEMD is an image decomposition tool that convoles no base function or filter, and separates a signal into functions characterizing each modes inside the signal, namely intrinsic mode functions(IMFs). In this study, a reasonable decomposition on fabric images is accomplished by optimizing the BEMD algorithms, which could generate an IMF1 containing pure fabric texture information, IMF2 and IMF3 involving intensity variety at lager scales. Different signal processing and segmentation methods were designed according to the information contained in each IMF and combined into an integral detection framework, including intensity detection and texture detection channals, to achieve segmentation on multiple kinds of fabric defect. The experiment results demonstrate that the proposed algorithm has a sound self-adaptivity and robustness. The main research contents and conclusions are as follows:(1) Through the analysis of human recognition on intensity changing and texture changing defects, and the self-adaptive decomposition ability of BEMD on defective fabric images, the theoretical evidences of fabric defect detection algorithm based on BEMD was proposed.A person will focus on the mutation of fabric surface color when he tries to visually recognize intensity changing defects, and will focus on the mutation of texture modes when recognizing texture changing defects. Usually the two kinds of recognition are performed concurrently, nonetheless, they can not be confused. The commonly used image decomposition tools in relating researches, such as wavelet decomposition, were not able to decompose a fabric image into texure signal and large-scale intensity changing signal according to the own features of the fabric image. Therefore, this thesis gave an analysis on BEMD theory and the capabilities of BEMD for fabric images’ self-adaptive decomposition, and consequently adopted BEMD as the core part of fabric defect detection algorithm.(2) By analyzing the effects of each procedure in BEMD and the problems exiting in relating researches, corresponding optimizations and improvements were proposed to achieve a reasonable decomposition on fabric images.First the key problems in BEMD were summarized into extrema finding, boundary processing, interpolation method selecting and design of stopping criterion. For the extrema finding, regional extrema was defined to including all kinds of etrema in 2D signas, and the morphologically geodesic dilation operator was used to find all the regional extrema. For the boundary processing, a mirroring extending method taking the image boundaries as axes was presented to significantly restrain the boundary effects. Through the comparing analysis on different RBF based and DT based interpolation methods, an interpolation scheme including DT- cubic interpolation, down sampling and RBF-thinplate smoothing was presented to give both fast compuatation speed and sound smoothness. Fot the design of stopping criterion, after the effects of stopping criterion and decomposition performance of actual fabric images were analyzed, the stopping criterion of SDMAX=0.2 was set. Experiment results showed that, with the optimizations of all the procedures in the BEMD, the fabric images could be saperated reseaonably and fast.(3) Based on the improved BEMD, a fabric defect detection algorithm including intensity detection channel and texture detection channel was designed.In the intensity detection channel, IMF2 and IMF3 were fused, to output a signal named “IMF2+3” which involve the intensity changing information as the input for segmentation. In the texture detection channel, total Hilbert transform and Laws texture energy measurement were used to extract the texture characters from IMF1 to provide input for segmentation. The segmentation’s input signals in both channels were checked with a single and double Gaussian fit, and were categorized into small-area defect(including defect-free samples) and large-area ones, they were then denoised with a hybrid fourth partial differential equation method, and finally segmented using confidence interval double thresholds or OTSU threshold, respectively.The final defect segmentation results can be acquired by merging the outputs from intensity detection channel and texture detection channel. In the off-line experiments, the detection algorithm correctly segmented 92.69% of the samples. Meanwhile, it was found that the great majority of defects were the mixture of two defect categories, so one single detection channel could not output complete result of defect area, and the combined detection scheme could generate better defect segmentation.(4) In order to improve the detection performance on purely texture-changing defects, the monogenic wavelet analysis was introduced as the new texture feature extraction tool to replace the total Hilbert transform and the Laws texture energy measurement.The experiments on IMF1 singals of fabric images demonstrated that the monogenic wavelet analysis is isotropic and has a sound response on texture orientation, and the tiny texture change in defect area could be responed in the orientation or amplitude signal at a resolution.For the orientation and amplitude signals at multiple resolutions, different criteria were design to select the best reponse. The orientation signals’ internal standard deviations were calculated as σ after phase shifted. If an orientation signal was with minimum σ and its σ was less than 0.1, it was identified as the best response. For the amplitude signals, design a parameter called max inter-class mean differential(MICMD) as the differential between means of foreground and background provided by OTSU segmentation. The amplitude signal with the largest MICMD was identified as the best response. A signal with best response was segmented with the segmentation suite represented in chapter four, to output defect detection result. The experiment results demonstrated that replacing Hilbert transform and Laws texture energy measurement with monogenic wavelet analysis could significantly improve the detection rate on texture changing defects, and the detection rate on overall experimental samples was increase to 98.93%.