Fusion-features and visual-dictionary image recognition methods for apple classification in smart manufacturing / Ahsiah Ismail
Smart manufacturing enables an efficient manufacturing process to optimize production. The optimization is performed through data analytics that requires reliable and informative data as input. Therefore, in this research, two image recognition feature extraction methods namely Curvelet Wavelet-G...
| Main Author: | |
|---|---|
| Format: | Thesis |
| Published: |
2020
|
| Subjects: | |
| Online Access: | http://studentsrepo.um.edu.my/14388/ http://studentsrepo.um.edu.my/14388/2/Ahsiah.pdf http://studentsrepo.um.edu.my/14388/1/Ahsiah_Ismail.pdf |
| Summary: | Smart manufacturing enables an efficient manufacturing process to optimize
production. The optimization is performed through data analytics that requires reliable
and informative data as input. Therefore, in this research, two image recognition feature
extraction methods namely Curvelet Wavelet-Gray Level Co-occurrence Matrix (CWGLCM)
and Fuzzy-Spatial Pyramid Matching (F-SPM) are proposed to provide reliable
inputs for vision-based apple classification in smart manufacturing. Feature extraction is
one of the major steps that could influent the efficiency of the manufacturing process.
The CW-GLCM method is a feature extraction of fusion-features with Decision Tree
classifier, while the F-SPM method uses a visual-dictionary based method to extract
features of visual pattern and the output is process by Support Vector Machine (SVM)
classifier. To evaluate the performance of the proposed methods, they are compared with
five existing methods, which are Bag of Words (BOW), Spatial Pyramid Matching
(SPM), Gray Level Co-occurrence Matrix (GLCM) Texture analysis, Convolutional
Neural Network (CNN) and Contrast‐Limited Adaptive Histogram Equalization + GLCM
+ Extreme Learning Machine (CLAHE+GLCM+ELM). Three datasets which are NDDA,
NDDAW and DA datasets with a total of 1310 apple images are collected to test the
proposed methods. The NDDA and NDDAW datasets are both binary-class of defective
and non-defective apple dataset, with NDDAW contains more low-quality region images
compared to the NDDA. Conversely, the DA dataset comprised of five different types of
defective apples to be used in multi-class tests. The proposed methods are trained and
evaluated using 10-fold cross-validation. Their classification accuracy, precision and
recall rate are then measured. Training and testing times are also recorded. From the
evaluation, the proposed F-SPM method attained 98.15% classification accuracy, 96.30% precision and 100% recall for NDDA, 91.07% for accuracy, 100% precision and 84.85%
recall for NDDAW, 86.33% for accuracy, 91.43% precision and 85.00% recall for DA
dataset. The F-SPM method outperformed the existing methods especially for NDDAW
and DA datasets. Alternatively, the CW-GLCM method able to obtain 98.15% accuracy,
96.30% precision and 100% recall for NDDA, 89.11% accuracy, 86.79% precision and
91.01% recall for NDDAW, 85.20% of accuracy, 88.33% precision and 85.00% recall for
DA dataset. The proposed CW-GLCM also shows the highest percentage (100%) for all
measurements (accuracy, precision and recall) and it even outperform others in
recognizing the Bruise defect. These results indicate that both proposed methods are
reliable and have the potential to be used for vision classification in smart manufacturing.
|
|---|