Classification and Compression Based on Component Analysis of Images
The application of visual pattern recognition in agriculture is a growing field of study. Accurate objective inspection and grading of agriculture products using a computer vision system hold great potential for product quality control and sorting. In this thesis, colour features of barley seeds are used as feature inputs to a multi-layer neural network in barley seed variety classification experiments. Four barley varieties, Tankard, Breedn, Creme, and CDC Dolly, were tested. 140 samples for each variety were used for pattern recognition training and an additional 140 samples for testing. Different colour spaces RGB, Lab, and HSI were used to represent colour features. The results showed that HSI colour features were the best for variety classification. The best recognition accuracy for individual seeds was 82.7% using a neural network with a structure of 4 inputs, 20 neurons in each of two hidden layers, and 4 outputs. With the increasing need to store and electronically transmit images, the study of image coding is getting more attention. A major objective of lossless image coding is to represent an image with as few bits as possible without loss of any information in the image. In this thesis, the statistical properties of an image were studied to compare lossless image data compression schemes. Huffman coding is reviewed, and two methods of constructing a difference image were implemented and evaluated. After applying these two methods to MRI images, compression ratios were calculated. The standard difference image had the lowest compression ratio. The advantage of an hierarchy embedded differential image is that the image quality can be progressively improved as the image transmission progresses. A coding method which considered the correlation of the adjacent pixels was also studied. The compression results by using this method on MRI images were discussed.