Natural Produce Classification Using Computer Vision Based on Statistical Color Features and Derivative of Radius Function .pptx

Natural Produce Classification Using Computer Vision Based on Statistical Color Features and Derivative of Radius Function .pptx

1. Natural Produce Classification Using Computer Vision Based on Statistical Color Features and Derivative of Radius Function

2. Outline Introduction Material and methods Computer vision system Image acquisition Segmentation Feature extraction Classification Validation Result

3. Introduction… Natural produce classification aims to classify the produce into one of classes . Applications in agriculture industry: Sorting Grading Measuring Pricing Traditionally is performed by human expert: Inaccurate Difficult to standardize Computer vision over promising method for natural produce classification.

4. Introduction… Current research in computer vision for natural produce classification: Using combination of very long features descriptor Unser’s descriptors Color coherence vectors Border/interior Global color histogram Appearance descriptors Color autocorrelogram Local activity spectrum Quantized compound change histogram Edge orientation autocorrelogram

5. Introduction Using complex classification method to obtain classification accuracy more than 97% Fusion of SVM with RBF kernel Bagging ensemble of linear discriminant analysis Fusion of binary classifier More time is required To extract the long features from an image To train complex classifier using the long features

6. Material and methods… Computer vision system: Camera: Logitech HD Pro Webcam C910 Light source: tube lamp light Personal computer: 3.00 GHz Pentium (R) Dual-Core, 2.00 GB RAM, 32-bit Windows 7 Software: coded in Matlab R2010a Produce Light source Computer Camera USB cable

7. Material and methods… Image acquisition: From top view The produce was placed on a table about 40 cm below the camera Acquired with a black background. Image: saved in RGB, 640 × 480 pixels, 96 dpi.

8. Material and methods… Segmentation To separate the produce from its background Converted to HSV Used only image in S channel Gaussian filter Automatic thresholding Morphological openings and closings (a) original image, (b) image in H channel, (c) image in S channel, (d) image in V channel, (e) binary image, (f) segmented image

9. Material and methods… Feature extraction: Statistical color (12 features) Mean Standard deviation Skewness Kurtosis Derivative of radius function (4 features) Mean Standard deviation Skewness Kurtosis Pixel values in H, S, and V channel  r (  ) 36 values of radius function is measured for with interval .  

10. Material and methods… Classification k -nearest neighbors (k-NN ) k was determined by experiment Euclidean distance was used to measure the closeness between testing and training sample Artificial neural network (ANN) : Three layers Input: 16 neurons Hidden: the number of neurons was determined by experiment Output : 3 neurons Transfer function: tansig Learning algorithm: Levenberg -Marquardt backpropagation

11. Material and methods Validation For preliminary experiment: 160 images of natural produce 53 apples 52 mangos 55 tomatoes 50 % samples were randomly chosen as training set and the rest for testing set 10 training sets and 10 testing sets were constructed Classification accuracy:

12. Result… k -NN classifier The number of nearest neighbors k was heuristically determined for The best average classification accuracy was achieved using and training : 94.25 % testing : 74.5%  

13. Result… ANN classifier The number of neurons in hidden layer was heuristically determined from two until 12 neurons The best average classification accuracy was achieved using 9 neurons in hidden layer training : 100 % Testing: 99.875%

14. Result Average processing time: Process a raw sample image and extract features: less than 0.22 second Classify testing set using k -NN: less than 0.01 second /sample Train ANN: less than 1 second Classify testing set using ANN : less than 0.01 second /sample Total time used to classify a new produce image: less than 0.23 second

15. Thank you