Image is everything - real-time leather defect detection systems

15 March 2017



Grading leather is crucial to determining its value, but the process is painstaking, labour-intensive and subject to human error. The Central Leather Research Institute’s Malathy Jawahar explains how intelligent, real-time leather defect detection systems can be achieved using image processing techniques based on artificial neural networks.


Inspecting quality is very important in assessing leather’s effective cutting value. As a natural material, its quality varies due to inherent variation. The price of the skin or leather varies considerably with the selection, so grading is done with great care and by experienced operators.

Manual inspections are highly subjective, though, and any differences in opinion can lead to disputes between buyers and sellers. It is also tediously repetitive, and fatigue can lead to defects passing unnoticed, leading to inaccuracy and inconsistency.

Intelligent automatic leather defect detection using image processing techniques is a possible solution to these problems. Optimal texture features – such as entropy, energy, contrast, correlation, cluster prominence standard deviation, mean, and local homogeneity – that discriminate between normal and defective leathers are extracted. Normal and defective leathers are classified using an artificial neural network features. Experimental results on a leather defect image library database suggest that the identification of defects can be automated using image processing based on feature-extraction techniques that use artificial neural networks (ANNs).

Consumer goods are increasingly under scrutiny. The global market places higher demands on quality, which means that quality control in leather production has never been more important and some form of objective quality assessment could play a vital role in meeting user requirements.

Visibly flawed

Establishing cutting value during manufacturing is based on the number and location of defects on the leather substrate. Because of this, the price of the leather varies considerably, which is why experienced assorters are necessary. There is no universally accepted grading system to describe surface defects, and each buyer or seller sets their own rules, which is how disputes can arise.

Computer-aided analysis provides a way for automatic surface inspection and defect identification. Image scrutiny and pattern recognition using mathematical algorithms based on texture have been employed to understand the nature of defects captured by the industrial CCD cameras.

Texture analysis of the images gives an overall perspective of the spatial distribution of pixels. Image texture gives information about the spatial arrangement of intensities in an image. It works in a way that is similar to visual perception in humans and handles uncertainties very effectively. Texture analysis includes collective analysis of local pixel regions, rather than relying on spatial information from single pixels. The goal is an intelligent, real-time defect detection system that is able to give quantitative descriptions of leather surface images for classification.

Wavelet transform

Wavelet transform is very useful for detection and matching applications. Multiresolution analysis is one of its important features. The application of repetitive low and high-pass filters gives the coefficients in four quadrants, such as approximation (low/low frequency), horizontal (low/high frequency), vertical (high/low frequency) and diagonal (high/high frequency).

Several bases are available. The entropy-based method has been used to select wavelets. For optimal wavelet selection, information extraction criteria and distribution error criteria are used. Based on the property of signal/image, wavelets were selected. There is no unique technique or parameters to choose the wavelet. Survey on mother wavelet selection method gives the broader view about the wavelet selection method.

The property of discrete, orthogonality and compact support properties were chosen to narrow down the selection of wavelets into four wavelets out of 15 wavelets. The four selected wavelets were Haar, Daubechies, Symlet and Coiflet. The level of decomposition gives an enhanced conception of an image. Based on the energy of the sub-bands, the level of decomposition is stopped. Two-level decomposition was done and further decomposition was not required, because the energy content became zero, and hence two-level decomposition was carried out.

Manual inspections are highly subjective… and any differences in opinion can lead to disputes between buyers and sellers. It is also tediously repetitive, and fatigue can lead to defects passing unnoticed, leading to inaccuracy and inconsistency.

Intelligent, real-time leather defect detection systems are designed to identify surface defects automatically. The leather imaging set-up consists of a feed and delivery rollers and a CCD camera attached to scanning carriage. As leather is conveyed by the rollers, a scanning carriage moves the camera horizontally, capturing ten images in each row (the number of rows varies between ten and 12). The size of each acquired image is around 1,600×1,200 pixels. The database contains a representative set of defective and non-defective samples that are used for training and testing.

Image Processing Toolbox technology was used to capture the whole leather surface: the images aid real-time monitoring of the defective and non-defective regions (see figure 1). To speed up computation time, the images were converted to greyscale and transformed to wavelet domain.

Wavelet analysis of leather images provides vital information about textural features reflecting the coarseness, smoothness, contrast and randomness of pixel distribution. Statistical measurements, such as higher-order moments and correlation represent similarity of pixels.

The feature set comprising of mean, standard deviation (SD), skewness, kutosis, homogeneity, inverse difference momentum (IDM), mean energy, mean absolute deviation (MAD), co-variance (CV), inter-quartile range (IQR), angular radial partitioning (ARP) and transform domain features act as predictor coefficients to identify the leather surface defects.

Epoch-making technology

Good leather surfaces are very smooth and homogeneous. This was evident with very low variance, in SD, MAD, CV and IQR. Similarly randomness was markedly reduced with lower entropy values. Defects – such as bacterial infection, fungal attack, grain damage, pox marking and lime blasting – showed porous patches with rough and coarse texture.

This led to high SD, MAD and CV values and randomness, with high entropy value. Defects such as chrome patch, dye patch and so on appear as smooth textures with variable intensity patches.

The feature values reflected these defects with higher mean intensity, probability distribution (kurtosis), skewness values and homogeneity.

The extracted discriminant texture features values were given as input to the ANN to classify the leather surface images as defective and non-defective (see figure 2). Initially, with known input-output pairing, the network was trained for 10,000 epochs. The network converged after 8,552 epochs with an error threshold of 0.118, underlining the accuracy of the test data. After training, the network was used to identify the defective and non-defective regions of the unknown whole leather images. Classification accuracy was above 90%, with chrome and dye patch defects showing close margins.

Using image processing algorithms and neural networks for identifying leather defects appears to be a promising means of achieving automatic inspection, and has shown the technical feasibility of an intelligent real-time inspection system that would eliminate human error. Automated inspection systems can accurately identify the surface defects to provide a consistent quality inspection, thereby increasing productivity and reliability.  ?

If you would ilke any further information about this article, please contact Leather International editor Carl Friedmann at [email protected]

 

Figure 1. Workflow diagram for leather surface defect identification.
Figure 2. ANN training plot.


Privacy Policy
We have updated our privacy policy. In the latest update it explains what cookies are and how we use them on our site. To learn more about cookies and their benefits, please view our privacy policy. Please be aware that parts of this site will not function correctly if you disable cookies. By continuing to use this site, you consent to our use of cookies in accordance with our privacy policy unless you have disabled them.