Sage Journals: Discover world-class research

Abstract

An auto fabric defect detection system via computer vision is used to replace manual inspection. In this paper, we propose a hardware accelerated algorithm based on a small-scale over-completed dictionary (SSOCD) via sparse coding (SC) method, which is realized on a parallel hardware platform (TMS320C6678). In order to reduce computation, the image patches projections in the training SSOCD are taken as features and the proposed features are more robust, and exhibit obvious advantages in detection results and computational cost. Furthermore, we introduce detection ratio and false ratio in order to measure the performance and reliability of the hardware accelerated algorithm. The experiments show that the proposed algorithm can run with high parallel efficiency and that the detection speed meets the real-time requirements of industrial inspection.

Keywords

Fabric Defect Detection Sparse Coding Hardware Acceleration

1. Introduction

With the rapid development of image recognition and hardware platforms, numerous fabric defect detection algorithms have been proposed in recent years. One of these is the grey-level co-occurrence matrix (GLCM) [1, 2, 3], which describes the second-order texture feature of images. However, the GLCM is unable to distinguish defects from images properly, due to its excessive matrix dimensions. Furthermore, some multi-resolution (multi-scale) analysis methods such as adaptive wavelet transforms (WT) [4] and Gabor transform [5] have been introduced to detect local defects by providing local resolution in horizontal, vertical and diagonal directions. In [7], independent components of fabric with normal textures that were different from the independent components of defective textures were obtained by resorting to independent component analysis (ICA), while the realization of the detection was based on the difference between the independent components. In addition, the combination of WT and ICA [8], Markov random fields [9] and SC [10, 11] were also applied to fabric defect detection and achieved good detection performance. However, many challenges remain such as the limits of computation and the requirements of detection ratio, which should be overcome when applying these algorithms in real industrial inspections. In this paper, we propose a fabric defect detection algorithm based on improved SC that can train the SSOCD; additionally, image patches' projection in the training SSOCD is taken as detection features, as these have obvious advantages for detection results and computational cost. Prior to the training process, the Gabor filter is used to preprocess fabric images based on classic frequency-domain defect detection algorithms, which can reduce the complexity of reconstruction and the influence of noise. The SSOCD is trained after completion of the above preprocessing.

Multi-core parallel hardware architecture is more suitable for computing intensive applications compared to conventional processors. TMS320C6678, which is a type of digital signal processor (DSP) with eight cores and integrated on a chip was recently developed by Texas Instruments. We employed this signal processor due to its outstanding fixed/float processing performance, optimized KeyStone multi-core architecture, flexible expandability and its capability for drawing from the strengths of eight cores to improve the overall performance. In this paper, we propose a parallel acceleration algorithm, based on a feature split for DSP to accelerate SSOCD extraction. Meanwhile, the detection system optimizes the type of synchronization and communication between different cores, which results in a reduction of the algorithm's serial time-usage, enabling the system to reach a state with a good parallel speedup ratio and parallel efficiency. Moreover, by optimizing the DSP program to take full advantage of parallel processing performance, the algorithm's processing time is greatly reduced. Finally, our analysis shows that the proposed algorithm can run with high parallel efficiency and the detection speed meets the requirements of industrial inspection.

In a real industrial system, cameras capture images from the fabric production channel and a fabric defect detection model automatically recognizes and alerts with regard to fabric defects. In order to achieve auto fabric defect detection via computer vision, our detection application is summarized as (1) acquiring fabric images in an industrial environment; (2) training the SSOCD via SC using a serious of fabric images without defects; (3) computing the features of fabric images to be detected; (4) judging and recognizing the defect.

The rest of this paper is structured as follows. We first describe fabric defect detection flow via SC and provide an overview of the SSOCD extraction in Section 2, then introduce a parallel hardware acceleration method for defect detection in Section 3. Section 4 analyses the experimental results and shows that our algorithm has superior robustness. We conclude this paper in Section 5 from a viewpoint of industrial application.

2. SSOCD of SC for Fabric Defect Detection Overview

The main problem in traditional fabric defect detection methods based on WT is deficiencies in standardization. Thus, a serious of training methods such as ICA [7, 8] and SC [10,11], which have the ability to flexibly receive effective descriptors by means of training fabric samples are proposed to detect fabric texture features. In this paper, we chose SC for our fabric defect detection, because its efficient feature extraction scheme is more suitable for parallel hardware acceleration.

In the scenario of fabric defect detection via SC, current hardware levels find it difficult to meet the demand for industrial real-time detecting, due to the significant computational resources needed for acquiring sparse expression. We therefore propose SSOCD for defect detection concerning the algorithm. In order to do so, three major steps need to be completed. Below, we provide a brief explanation of the flow of fabric defect detection via SSOCD.

2.1. Training SSOCD of SC

The process of training SSOCD of SC simulates the human visual cortex in terms of information processing [12]. And SSOCD of SC reflects to some degree the most basic elements of an image. Assume that $X = [x_{1}, x_{2}, \dots, x_{M}]^{T} \in R^{M \times N}$ represents the image patches of twill without defects, where M is the number of image patches and N represents the dimensions of image patches. Training the SC of an image can then be written as the following formula:

\begin{array}{l} C = \underset{C, S}{argmin} {‖ X - S C ‖}^{2} + λ | S | \\ s . t . ‖ c_{k} ‖ \leq 1, \forall k =1,2, …, K \end{array}

(1)

where $S = [s_{1}, s_{2}, \dots, s_{M}]^{T} \in R^{M \times K}$ represents the sparse expression of image patches and s_m, which is the m'th row vector of S, represents the corresponding sparse expression of x_m, which is the m'th row vector of X, $C = [c_{1}, c_{2}, \dots, c_{K}]^{T} \in R^{K \times N}$ is the SSOCD of SC, $λ | S |$ represents the sparse penalty function that measures the sparse degree of coding, K is the scale of SSOCD and c_k is the k'th row vector of C. Usually, images are preprocessed using a Gabor filter to decrease the scale of the dictionary, due to normal twill having strong periodicity and consistency. Fig.1 indicates the processing flow for training SSOCD of SC. For an original fabric image with a size of H1×W1 to be trained in our image database, it is preprocessed by a Gabor filter to obtain a a new fabric image whose size is H2×W2 at first. Secondly, we extract image patches with a size of $h \times w$ by scanning rows and columns one by one; in this way, we can get $(H 2 - h + 1) * (W 2 - w + 1)$ image patches from this trained image. Finally, M image patches are extracted from our image database and trained to acquire SSOCD of SC.

Figure 1.

Overview of SSOCD training

2.2. Feature Extraction

Feature selection is an important part of the detection process [10,11] used to reconstruct errors and is achieved by calculating coding coefficients as features to detect defects once the dictionary has been trained. The reconstruction error is calculated as follows:

\begin{array}{l} S_{d}^{*} = \underset{S_{d}}{argmin} {‖ X_{d} - S_{d} C ‖}^{2} + λ | S_{d} | \\ X_{d} = S_{d}^{*} C \\ E r r o r F_{d} = {‖ X_{d} - {\hat{X}}_{d} ‖}^{2} \end{array}

(2)

where $X_{d} \in R^{1 \times N}$ represents an image patch, S_d is its corresponding sparse expression, ${_{X}}_{d}$ is the reconstructed image patch and $E r r o r F_{d}$ is the reconstruction error of the image patch to be detected. The greater the $E r r o r F_{d}$ , the greater the likelihood of detection; however, the fabric image likely cannot be reconstructed with a small $E r r o r F_{d}$ and will require significant computation. So $E r r o r F_{d}$ is not a suitable feature for our detection.

We introduce every patch's projection in the training SSOCD as the detection feature of an image. For an image patch X_d, its feature $F_{d} = [f_{1 d}, f_{2 d},…, f_{K d}]^{T} \in R^{K \times 1}$ is computed by:

F_{d} = C X_{d} ​^{T}

(3)

The process of extracting projection features is shown in Fig.2. We can see that the process of preprocessing a fabric image to be detected is similar to preprocessing in the training process and the last projection feature of this fabric image, $F = [F_{1}, F_{2}, \dots, F_{D}] \in R^{K \times D}$ , which can also be written as $F = [{\bar{F}}_{1}, {\bar{F}}_{2}, \dots, {\bar{F}}_{K}]^{T} \in R^{K \times D}$ , can be calculated by Eq. 3. D, which represents the amount of image patches that are extracted from the image to be detected.

Figure 2.

The process of feature extraction via projection

Figure 3.

The process of feature extraction via projection

Sparse expression can be computed in a situation where there is little reconstruction error from fabric with a normal texture, because SSOCD is also trained from fabric with normal texture. For a SSOCD vector, if its corresponding sparse expression value is not zero, the value of normal image patch projection in this vector should be large and the value of an image patch with defects should be small. In other words, although the projection feature F_d of X_d is not precisely equal to sparse expression $S_{d}^{*}$ , F_d, it nonetheless roughly represents the variation trend of $S_{d}^{*}$ . In this regard, setting the projection feature as a detection standard is reasonable. For the purpose of clearly distinguishing this feature, we improved the projection feature of X_d, accounted for by Eq. 4. One projection feature is the sum of the absolute value of projection, the other one is the largest absolute value of projection and they are defined below as follows:

\begin{array}{l} S u m F_{d} = \sum_{k = 1}^{K} | f_{k d} | \\ M a x F_{d} = \max {| f_{1 d} |, | f_{2 d} |,…, | f_{K d} |} \end{array}

(4)

Fig.3 shows that these three features ( $S u m F_{d}$ $M a x F_{d}$ $E r r o r F_{d}$ ) have different degrees of distinction for fabric images. The feature images of $(c)$ are constructed from image patches where every image patch ${X^{'}}_{d}$ is calculated by Eq. 5, while image patches where every image patch $X ″_{d}$

\begin{array}{l} {X^{'}}_{d} = X_{d} - {F^{'}}_{d} ​^{T} C \\ {F^{'}}_{d} = {[| f_{1 d} |, | f_{2 d} |,…, | f_{K d} |]}^{T} \end{array}

(5)

is calculated by Eq. 6 constructs the feature images of $(d)$ . We can see that all features have a good degree of differentiation between fabric with defects and fabric without defects in the first column images; however, for the last two column images, features via the value of projection ( $S u m F_{d}$ $M a x F_{d}$ ) have better robustness compared to the reconstruction error $(E r r o r F_{d})$ from Fig.3.

\begin{array}{l} X ″_{d} = X_{d} - F ″_{d} ​^{T} C \\ F ″_{d} = {[f ″_{1 d}, f ″_{2 d},…, f ″_{K d}]}^{T} \\ f ″_{k d} = {\begin{array}{l} M a x F_{d}, | f_{k d} | = M a x F_{d} \\ 0, | f_{k d} | \neq M a x F_{d} \end{array} \end{array}

(6)

2.3. Fabric Detection

In this paper, the average $A = [a_{1}, a_{2},…, a_{K}]^{T} \in R^{K \times 1}$ calculated from trained samples is used as the detection standard and is given by:

a_{k} = \frac{1}{M} \sum_{m = 1}^{M} f_{k m}, k = 1,2,…, K

(7)

In Eq. 7, M represents the amount of trained image patches from our image database and $f_{k m}$ represents the projection in a situation where the $m^{'} t h$ image patch is projected to the $k^{'} t h$ coordinate of the SSOCD.

\begin{array}{l} V = \sum_{k = 1}^{K} {({\bar{F}}_{k} \otimes H_{a v g} - a_{k})}^{2} \\ R = {\begin{array}{l} 1, V ≻ T \\ 0, V ≺ T \end{array} \end{array}

(8)

Eq. 8 decides whether an image is recognized as a defect image or not, where V is the distance between the image to be detected and the trained image, ${\bar{F}}_{k}$ is the $k^{'} t h$ row vector of F, as explained in the process of extracting the projection feature above, $H_{a v g}$ represents the mean filter and is introduced for the purpose of improving detection accuracy and robustness, and R is the result of detection. Fig.4 shows the process of detection.

Figure 4.

The process of detection based on feature

3. Parallel Hardware Acceleration for Fabric Defect Detection via SC

TMS320C6678 with an eight-core DSP has a suitable hardware architecture for parallel acceleration of fabric detection. Fig.5 summarizes the actual data processing procedure of fabric defect detection via SSOCD. Three important modules are shown in Fig.5, i.e., Gabor filtering preprocessing, feature extraction and mean filtering processing; these can be converted to the operation of image template convolution. The final size of the Gabor filtering convolution template is 5*5, the mean filtering convolution template is 16*16 and the final scale K and dimension N of SSOCD is 16 and 30, respectively. We propose a parallel acceleration architecture, which is based on feature splits according to the characteristics of detection algorithms.

Figure 5.

The parallel acceleration flow of defect detection

3.1. Hardware Acceleration Architecture Based on Feature Split

The processing flow of this architecture is associated with the allocation of eight cores. As shown in Fig.6, we can see that parallel acceleration flow is based on a feature split. First, when the Gabor filter preprocesses an image acquired from an industry camera, the image to be detected is divided into eight equal image blocks and each image block is assigned to a core for processing. For each image block, it is necessary to keep overlapping pixels on the edge of an image block for the purpose of eliminating the impact of invalid data produced by convolution filtering. Additionally, the eight preprocessed image blocks are synchronously assembled into one complete image. Furthermore, the procedure of feature extraction from the detection image is the core step and K, which represents the number of SSOCD, determines the dimensions of the image features and the detection quality; therefore, we propose the architecture via a feature split in order to accelerate detection by assigning the procedure of K dimensional feature extraction to each core. Finally, a result can be obtained by judging the distance between detected images and trained images.

Figure 6.

The parallel acceleration flow of defect detection

3.2. Hardware Acceleration at the Code Level

For fabric defect detection via SSOCD, computation primarily depends on mask convolution. The essence of mask convolution lies in multiply-add operation between mask coefficients and fabric images to be detected. If we use the two-dimensional convolution method to calculate mask convolution, it causes low efficiency of compilation, a long time using registers and low parallelism of command and data. Particularly when the column of an image increases, row address of image data has a wide range of jump and decreases the read/write hit ratio of caching, resulting in an increase of CPU resources required.

To decrease time consumption, a one-dimensional local convolution method that can improve the efficiency of compilation is introduced. When using, for example, a Gabor convolution with a 5*5 template for a pixel in a certain row of an image, its local convolution result is calculated by multiplying 5 pixels' grey value over the pixel with their corresponding Gabor convolution template coefficients; we can then get the local convolution results of a specific row and add them to a cache array. Next, the same operation is executed four times and we can get the Gabor convolution results of a certain row from the cache array. This method effectively accelerates our algorithm on the hardware platform as a result of increasing the read/write hit ratio of the cache.

4. Experimental Results Analysis

During the procedure of fabric detection, we applied the TILDA public database of twill as test images. In this section, we discuss the relationship between the detection ratio and algorithm parameters. Following on, we compare our proposed method with other methods using the detection ratio as an evaluation criterion. Finally, an analysis of hardware acceleration is provided.

4.1. Relationship between Detection Ratio and Algorithm Parameters

The size of image patches is an important parameter in our algorithm; if it is too small, the detection result will easily be affected by noise through, for example, illumination changing, fabric grain fluctuating, etc. As shown in Fig.7, when an image patch that consists of 4*3 pixels is used as an element, the detection ratio is always below 90%, regardless of dictionary scale. Additionally, we can see that the size of image patches should be appropriate in case the scale of SSOCD remains fixed; the scale of SSOCD should also be appropriate in case image patch size remains fixed in order to obtain a high detection ratio and low false ratio (see Fig.7). Although large image patch size or SSOCD scale can improve detection ratio and decrease false ratio, it gives rise large calculations. A larger image patch size does not necessarily provide a better detection ratio. Detection ratio can exceed 90% in conditions where the scale of SSOCD is greater than 24 and when the image patch consists of 12*10 pixels. However, image patches that consist of 6*5 or 9*7 pixels can also result in a detection ratio above 90% with a small dictionary scale, because 6*5 and 9*7 is near to the period of our experimental twill samples. Therefore, we can achieve an acceptable detection ratio by applying 6*5 or 9*7 as the size of image patches and 16 as the dictionary scale.

Figure 7.

Show of detection ratio and false ratio

The most suitable size for image patches was received as being near to the period of our fabric. Generally, we set the scale of the dictionary at slightly more than half the size of image patches. Detection results via SC are shown in Fig.8.

Figure 8.

Detection result of twill using SC

4.2. Comparison with Other Detection Methods

In our experiments based on the TILDA database, the method via SC was compared with the following methods in terms of detection performance: 1) ICA: the realization of fabric defect detection via ICA by replacing ICA with SC in the process of training and projection was tested, and the same dimension $(K = 16)$ and size of image patches (6*5) were used to make the comparison fair; 2) GLCM: this method considered that normal fabric texture has regularity and directional characteristics, and defects will destroy this rule. This method realized defect detection by describing comprehensive information regarding direction, adjacent interval and variation range in fabric grey image; 3) Gabor filter: this method calculates variance coefficients as features for detection after processing the fabric image using a Gabor filter. 4) WT: this method used a suitable WT coefficient to detect fabric defects. We tested these methods and calculated an average detection ratio (ADR) for different defect types according to hardware implementation. Fig.9 shows that the ADR of SC was found to be greater than for other methods. We applied one-way ANOVA for experimental data to reach our conclusion more persuasively; the F value of ANOVA is $7.23 > F_{0.05 (4,60)}$ , where 4 and 6 represent the degree of freedom of our experimental data; the P value of ANOVA is $8.02579 \times 10^{- 5} < 0.05$ , where 0.05 represents the inspection level. Therefore, the detection ratio of five groups showed statistic differences and the different methods were therefore shown to affect the efficiency of detection.

Figure 9.

Comparison of detection ratio

4.3. Parallel Hardware Accelerating for SC Detection Analysis

4.3.1. The Evaluation Criterion for Parallel Algorithm Performance

1. Speedup

An important indicator for measuring whether a parallel algorithm is a good fit is to compare a task's execution time in a single-core processor, with the processor having N parallel cores. The speedup of a processor having N parallel cores is defined as follows:

S (N) = \frac{T_{p} (1)}{T_{p} (N)}

(9)

where $T_{p} (1)$ represents the execution time of an algorithm in a single-core system, $T_{p} (N)$ represents the execution time of the same algorithm in a multi-core system. If we ignore the time of communication between different cores and memory, the above equation can be defined as:

S (N) = N

(10)

2. Parallel efficiency

Parallel efficiency is the ratio of speedup to the number of cores and can be defined as follows:

E (N) = \frac{S (N)}{N} \times 100 %

(11)

In an ideal situation, the maximum value of speedup is N and the maximum value of parallel efficiency can reach 1. We can evaluate the performance of the parallel algorithm and further improvement thereof from the above two evaluation standards.

4.3.2. Experimental Results and Analysis

In the TILDA public database of twill, the size of the image is 512*768, the convolution template of Gabor filtering is 5*5, the final scale K of SSOCD is 16 and the size of an image patch is 6*5. Firstly, a single-core serial detection algorithm is implemented using a Matlab inter core PC platform with 2 Duo CPU, E7300 @2.66GHz and 2.00GB memory. Furthermore, the detection algorithm is transplanted to the evaluation board of TMS320C6678 with single-core serial and multi-core parallel implementation and optimization. Finally, we analysed the performance of our detection algorithm via experimental results.

From Table 1, we can see that the execution time for the PC platform was 58 times as long as on the single core of TMS320C6678 for the performance of floating-point calculation; this was due to the application of the hardware acceleration algorithm and the powerful computing capacity of DSP. Additionally, the execution time of the single-core fixed-point calculation was less than single-core floating-point calculation on the TMS320C6678 platform.

Table 1.

Performance test on different platform environments

Platform environment	Floating-point calculation on PC	Floating-point calculation on a single core of TMS320C6678	Fixed-point calculation on a single core of TMS320C6678
Execution time(s)	65	1.112	0.713

As shown in Table 2, under the fixed-point environment of DSP, we tested consuming time by detecting a fabric image using different numbers of cores. Fig.10 shows that the speedup increased continually as more cores were added and parallel efficiency decreased as more cores were added. Due to bus competition between different cores, parallel efficiency and speedup could not reach their ideal maximum. When all eight cores were utilized, the average time consumption was 101.057ms, parallel efficiency was 88.18% and speedup was able to reach up to 7.054. Therefore, the algorithm presented in this paper is extremely suitable for parallel detection of fabric defects.

Table 2.

Consuming times for the different number of cores

Core number	Consuming time(ms)						Average consuming time
8	101.111	101.309	100.702	100.788	101.220	101.017	101.057
8	100.877	101.017	100.852	101.470	101.150	101.171	101.057
4	190.240	188.931	188.056	190.325	190.370	189.898	189.615
4	189.191	190.438	189.245	188.588	189.759	190.337	189.615
2	379.260	377.717	376.489	376.727	377.294	378.888	377.452
2	376.035	376.697	379.213	376.853	376.630	377.621	377.452
1	712.164	713.725	711.937	712.715	714.795	712.876	712.903
1	711.490	712.409	713.417	712.671	712.793	713.844	712.903

Figure 10.

The left plot shows the relationship between speedup and the number of cores. On the right is a plot for parallel efficiency for the different number of cores.

5. Conclusion

In this paper, we presented the detailed design and implementation of a parallel accelerating algorithm via SC for fabric defect detection. Firstly, a feature split was utilized to accelerate SSOCD extraction in order to meet real-time detection requirements. Secondly, by incorporating an optimized synchronization and communication method, which was able to reduce the algorithm's serial time-consuming, the system reached a state with a good parallel speedup ratio and parallel efficiency. Moreover, by optimizing the DSP program to make full use of parallel processing performance, the algorithm's processing time was significantly reduced. The experiments conducted show that the proposed algorithm can run with high parallel efficiency and that the detection speed meets the requirements of industrial inspection.

References

Haralick

R.M.

Shanmugam

Dinstein

I.H.

. Textural features for image classification. IEEE Transactions on Systems, Man and Cybernetics, SMC-3(6):610–621, Nov. 1973.

Benco

Hudec

Kamencay

Zachariasova

Matuska

. An Advanced Approach to Extraction of Colour Texture Features Based on GLCM. International Journal of Advanced Robotic Systems, Nov. 2014.

Raheja

J.L.

Kumar

Chaudhary

. Fabric defect detection based on GLCM and Gabor filter. Optik-International Journal for Light and Electron Optics, 124(23):6469–6474, Dec. 2013.

Yang

X.Z.

Pang

G.K.H.

Yung

N.H.C.

. Discriminative fabric defect detection using adaptive wavelets. Optical Engineering, 41(12):3116–3126, Dec. 2002.

Bodnarova

Bennamoun

Latham

. Optimal Gabor filters for textile flaw detection. Pattern recognition, 35(12):2973–2991, Dec. 2002.

Cho

Roberts

Jung

Choi

Moon

. An Efficient Hybrid Face Recognition Algorithm Using PCA and GABOR Wavelets. International Journal of Advanced Robotic Systems, 11(59):1–8, 2014.

Sezer

O.G.

Ertiizun

Erçil

. Independent component analysis for texture defect detection. Pattern Recognition and Image Analysis C/C of Raspoznavaniye Obrazov I Analiz Izobrazhenii, 14(2):303–307, 2004.

Serdaroglu

Ertuzun

Erçil

. Defect detection in textile fabric images using wavelet transforms and independent component analysis. Pattern Recognition and Image Analysis, 16(1):61–64, Jan. 2006.

Cohen

F.S.

Fan

Attali

. Automated inspection of textile fabrics using textural models. IEEE Transactions on Pattern Analysis and Machine Intelligence, 13(8):803–808, Aug. 1991.

10.

Zhou

Semenovich

Sowmya

Wang

. Sparse Dictionary Reconstruction for Textile Defect Detection. In IEEE 2012 11th International Conference on Machine Learning and Applications (ICMLA), Coll. of Textiles, Donghua Univ, Shanghai, China, pages 21–26, Dec. 2012.

11.

Zhou

Wang

. Fabric defect detection using adaptive dictionaries. Textile Research Journal, Oct. 2013.

12.

Olshausen

B.A.

. Emergence of simple-cell receptive field properties by learning a sparse code for natural images. Nature, 381(6583):607–609, 1996.

13.

Ferster

Chung

Wheat

. Orientation selectivity of thalamic input to simple cells of cat visual codex. Nature, 380(6571):249–252, 1996.

14.

Lee

Battle

Raina

A.Y.

. Efficient sparse coding algorithms. Advances in neural information processing systems, 801–808, 2006.

15.

Jing

Zhang

. Improved Gabor filters for textile defect detection. Procedia Engineering, 15:5010–5014, 2011.

16.

Wright

Yang

A.Y.

Ganesh

Sastry

S.S.

. Robust face recognition via sparse representation. IEEE Transactions on Pattern Analysis and Machine Intelligence, 31(2):210–227, Apr. 2009.

17.

Kuo

C.F.J.

Hsu

C.T.M.

Chen

W.H.

Chiu

C.H.

. Automatic detection system for printed fabric defects. Textile Research Journal, 82(6):591–601, 2012.

18.

Ngan

H.Y.T.

Pang

G.K.H.

Yung

N.H.C.

Automated fabric defect detectionąła review. Image and Vision Computing, 29(7):442–458, Jun. 2011.

19.

Rolls

E.T.

Tovee

M.J.

. Sparseness of the neuronal representation of stimuli in the primate temporal visual cortex. Journal of Neurophysiology, 73(2): 713–726, 1995.

20.

Young

M.P.

Yamane

. Sparse population coding of faces in the inferotemporal cortex. Science, 256(5061):1327–1331, 1992.