Fine-Grained Complex Image Classification Method Based on Butterfly Images

Fine-grained image classification is very difficult task and there is currently no effective machine learning method yet. Taking the butterfly images as an example, this paper comprehensively studies image classification methods for finegrained and complex images, focusing on improving the existing methods in the four aspects of image preprocessing, feature extraction, feature coding, and classifier design. An effective machine learning method for butterfly classification is established. In the aspect of image preprocessing, we firstly introduce the edge breakpoint connection method to make up for the defect that discontinuous edge can’t extract the target region effectively. In terms of feature extraction, nonseparable Shannon wavelet is used to combine with corner feature, blob feature, edge curvature feature and invariant moment feature selection to further improve the utilization of image information. In the aspect of feature encoding, most features in this paper are encoded by histogram, which increases the classification efficiency, and for histogram encoding features, an improved histogram intersection distance is proposed, which makes it more effective in KNN classifier. Finally, in the classifier design, the Bagging ensemble method is integrated into the parallel KNN classifier. Experiments show the effectiveness and robustness of the proposed method.