Vision based inspection system for leather surface defect detection using fast convergence particle swarm optimization ensemble classifier approach

Abstract

Surface defect inspection plays a vital role in leather manufacturing. Current practice involves an expert to inspect each piece of leather individually and detect defects manually. However, such a manual inspection is highly subjective and varies quite considerably from one assorter to another. Computer vision system for natural material like leather is a challenging research problem. This study describes the application of computer vision system to capture leather surface images and use of a novel Fast Convergence Particle Swarm Optimization (FCPSO) algorithm on a set of handcrafted texture features viz., GLCM and classified using supervised classifiers viz., Multi Layer Perceptron (MLP), Decision Tree (DT), SVM, Naïve Bayes, K-Nearest Neighbors (KNN) and Random Forest (RF). FCPSO using modified fitness function by selective band Shannon entropy is implemented to segment industrial leather images. Segmentation efficiency of the proposed FCPSO algorithm is evaluated and its performance is compared with other optimization algorithms. Efficiency of the segmentation algorithms is evaluated using performance measures such as average difference (AD), Area Error Rate (AER), Edge-based structural similarity index (ESSIM), F-Score, Normalized correlation coefficient (NK), Overlap Error (OE), structural content (SC), Structural similarity index (SSIM) and Zijdenbos similarity index (ZSI). Correlation of the segmented area using FCPSO with the experts’ ground truth is found to be high with R value of 0.84. Feature extraction is carried out using GLCM texture features and the most prominent features were selected using statistical t-test and correlation coefficient. Experimental results showed encouraging results for random forest classifier confirming the potential of the proposed system for automatic leather defect classification.

Appendix

1.1 Derivation of Haralick’s texture features from grey level co-occurrence matrix

P(i,j) is (i,j) th entry in normalized GLCM; P_x(i) is ith entry in the marginal probability matrix obtained by summing the rows of P(i,j); P_x(i)= ${\sum}_{\mathrm{j}=1}^{\mathrm{G}}\mathrm{P}\left(\mathrm{i},\mathrm{j}\right)$; P_y(j) is jth entry in the marginal probability matrix obtained by summing the columns of P(i,j); P_y(j)=${\sum}_{\mathrm{j}=1}^{\mathrm{G}}\mathrm{P}\left(\mathrm{i},\mathrm{j}\right)$; G is number of gray levels in quantized image. μ_x, μ_y,σ_iσ_j are the means and standard deviation of p_x and p_y, respectively.

No.	Feature name	Equation	Description
1	Autocorrelation	$\sum \limits_{i,j=1}^N\left(\mathsf{i}\mathsf{j}\right)\mathsf{P}\left(\mathsf{i},\mathsf{j}\right)$	Returns probability occurrence of the specific a pixel
2	Contrast	$\sum \limits_{i,j=1}^N\mathsf{P}\left(\mathsf{i},\mathsf{j}\right){\left(\mathsf{i},\mathsf{j}\right)}^2$	Returns the local variation boundary of the gray levels in a texture
3	Correlation	$\sum \limits_{i,j=1}^N\left(\mathsf{i}-{\mathsf{\mu}}_i\right)\left(\mathsf{i},{\mathsf{\mu}}_j\right)\mathsf{P}\left(\mathsf{i},\mathsf{j}\right)/{\mathsf{\sigma}}_i{\mathsf{\sigma}}_j$	Returns the probability occurrence of the specific pixel pair
4	Cluster prominence	$\sum \limits_{i,j=1}^N{\left(\mathsf{i}+\mathsf{j}-{\mathsf{\mu}}_x-{\mathsf{\mu}}_y\right)}^4\mathsf{P}\left(\mathsf{i},\mathsf{j}\right)$	Measures the skewness of the matrix (asymmetry)
5	Cluster shade	$\sum \limits_{i,j=1}^N{\left(\mathsf{i}+\mathsf{j}-{\mathsf{\mu}}_x-{\mathsf{\mu}}_y\right)}^3\mathsf{P}\left(i,\mathsf{j}\right)$	Measures the skewness of the GLCM (asymmetry)
6	Dissimilarity	$\sum \limits_{i,j=1}^N\mid i-j\mid P\left(i,\mathrm{j}\right)$	Returns linear local variation boundary of the gray levels in a texture
7	Energy (uniformity)	$\sum \limits_{i,j=1}^NP{\left(\mathsf{i},\mathsf{j}\right)}^2$	Returns the sum of squared elements in the GLCM.
8	Entropy	$-\sum \limits_{i,j=1}^NP\left(\mathsf{i},\mathsf{j}\right)\mathit{\log}\left(i,\mathsf{j}\right)$	Returns the degree of randomness and complexity of a texture
9	Homogeneity	$\sum \limits_{i,j=1}^N\frac{\mathsf{P}\left(i,\mathsf{j}\right)}{1+\mid \mathsf{i}-\mathsf{j}\mid }$	Returns the distribution consistency of the GLCM elements
10	Maximum probability	$ma{x}_{i,j}\mathsf{P}\left(i,\mathsf{j}\right)$	Returns the largest P(i,j) value found within the window
11	Sum of square (variance)	$\sum \limits_{i,j=1}^N{\left(\mathsf{i}-\mathsf{\mu}\right)}^2\mathsf{P}\left(\mathsf{i},\mathsf{j}\right)$	Measures the dispersion of the values around the mean
12	Sum average	$\sum \limits_{i=0}^{2G}\mathsf{i}{P}_{x+y}$	Measures the mean value of Px and Py
13	Sum variance	$\sum \limits_{i=1}^{2G}{\left(\mathsf{i}-\left(\mathsf{sum}\ \mathsf{entropy}\right)\right)}^2{P}_{x+y}(i)$	Measures the variance of sum entropy
14	Sum entropy	$\sum \limits_{i=0}^{2G}{P}_{x+y}(i)\mathsf{\log}\left({P}_{x+y}(i)\right)$	Measures the sum of Px and Pyentropy
15	Difference variance	VarianceP(x + y)	Measures the variance difference between two values
16	Difference entropy	$-\sum \limits_{i=0}^G{P}_{x-y}(i)\mathit{\log}\left({P}_{x-y}(i)\right)$	Measures the difference of Px and Pyentropy
17	Information measure of correlation (1)	$\frac{HXY- HXY1}{\mathsf{\max}\left( HX, HY\right)}$	Measures the linear dependency of gray levels on those of neighboring pixels.
18	Information measure of correlation (2)	${\left(1-\mathsf{\exp}\left[-2\left(\mathsf{HXY}2-\mathsf{HXY}\right)\right]\right)}^{0.5}$	Measures the linear dependency of gray levels on those of neighboring pixels.
19	Inverse difference normalized (INN)	$\sum \limits_{i,j=1}^G\frac{1}{1+{\left|i-j\right|}^2/{G}^2}P\left(i,j\right)$	Returns the linear normalized distribution consistency of the GLCM elements
20	Inverse difference moment normalized	$\sum \limits_{i,j=1}^G\frac{1}{1+{\left(i-j\right)}^2/{G}^2}P\left(i,j\right)$	Returns the normalized distribution consistency of the GLCM elements