Creating a robust image classification system depends on having enough data with which one can adequately train and validate the model. If there is not enough available data, this assumption may not hold and would result in a classifier that exhibits poor performance, thus lowering it's acceptability. This paper offers a solution to the problem of training and testing a neuro-fuzzy system for the purpose of image recognition when there are a limited number of images. Features of interest are segmented from each image and then used to train a neural-fuzzy system. This increases the number of data examples used to train the system. The neuro-fuzzy system is then tested on the entire data set set of full images. A high level of classification accuracy has been obtained using this method. This solution has two advantages; one, it overcomes the problem of limited data examples for training a classification model and two, rules can be extracted from the neuro-fuzzy model for further analysis. We apply this system to the problem to detection of pest damage on images of apples in New Zealand orchards.
|Cite as: Woodford, B.J., Deng, D. and Benwell, G.L. (2004). A Wavelet-Based Neuro-Fuzzy System for Data Mining Small Image Sets. In Proc. Australasian Workshop on Data Mining and Web Intelligence (DMWI2004), Dunedin, New Zealand. CRPIT, 32. Purvis, M., Ed. ACS. 139-143. |
(local if available)