Plant Leaf Image Classification By Using Persistent Homology And Multi-feature Fusion

Plant cultivar classification is essential to plant resource investigation and crop breeding research.As one of the most crucial plant organs,leaf is an important criterion to distinguish different plants.Therefore,studies about plant classification based on leaf image has been sheared on more sights and achieved notable results in recent years.However,most studies were about plant species recognition,and accurate plant cultivar recognition methods were still lacking.As a more fine-grained image classification task,the plant cultivar recognition faces the challenge of high leaf similarity between different cultivars when compared with species recognition.Therefore,finding out the more descriptive and discriminative features is the key problem to classify different cultivars.To overcome these obstacles,a leaf image classification method based on persistent homology and multi-feature fusion was proposed in this thesis,and the transparently scanned leaf images were taken as the research objects.The purposes of this thesis are to study the fine-grained leaf image classification and realize plant cultivar recognition.The main works of this thesis are listed as below:(1)Because the leaf inter-cultivar difference is not obvious while the intra-cultivar difference is existing,more leaf details should be retained in the feature extraction process to classify different cultivars.A persistent homology-based leaf topological feature extraction algorithm was proposed in this thesis,which extracted features from the high-definition leaf scanning images and retained more leaf details.The topological features of leaf shape,texture,and venation were extracted,respectively,and all of them were combined as the multi-perspectives topological features(CMTF).Finally,the integrated feature was used to recognize the leaf images from different cultivars.The experimental results showed that the multi-perspectives leaf topological features had a more vital discrimination ability than those extracted from the single perspective and achieved much higher accuracy than existed methods in several cultivar datasets.(2)To improve the accuracy of cultivar recognition,the topological feature-based classification method and Convolutional Neural Network(CNN)-based classification method were combined to fuse topological features and high-level image features extracted by CNN.The experimental results showed that higher accuracy than the existing methods had been achieved in several cultivar datasets after the fusion of these two features.The performance on the species datasets also indicated the considerable generalization ability of the proposed method.(3)Although the classification method that combined the topological features and high-level image features extracted by CNN had dramatically improved the accuracy,how the topological features influenced the classification was still not clear.To enhance the interpretability of the proposed method,we analyzed the changes of saliency map based on the CNN feature visualization technology.And we also analyzed the result of comparative experiments to evaluate the impact of topological features on accuracy improvement.The analysis results showed that the fusion of topological features made the network focus more on the characteristics of the leaf itself and promoted the learning ability of the CNN model.The main contributions of this thesis include proposing the leaf topological feature extraction method based on persistent homology and combining the topological features with the image features extracted by CNN to dramatically improve the accuracy of leaf image classification among different cultivars.The results demonstrated that the proposed method had a significant performance improvement in fine-grained leaf classification among cultivars and could also be applied to a broader range of leaf image classification among species.