The Study Of Plant Leaf Recognition Based On The Deep Belief Networks

Recognizing the leaf images in as short a time as possible has an important practical significance to protect and study plants, but manual analysis and processing the explosive growth of leaf images are almost impossible. Therefore, using computer-aided plant recognition to improve plant leaves recognition efficiency and reduce labor costs has become a hot research topic. However, when plant species increase to a large number, the leaf recognition rate becomes low since the traditional classification methods extract the few characteristics or use the classifier with simple structure. To solve this problem, we applied the multi feature fusion as a feature vector and used deep belief networks as a classifier for identification in this thesis. The main research content and conclusions are as follows:1. In this thesis, we used the typical Flavia dataset and ICL dataset. First of all, we made a preprocessing step for the leaf images, including image filtering, image rotating, and image cropping, to reduce interference of shooting problems. Secondly, the features were calculated, including Fourier descriptor, Gabor filter features, local binary pattern features, Hu invariant moments, gray level co-occurrence matrix features.2. Deep belief networks can create complicated structure, which we used as the classifier. And we used a "dropout" algorithm in training process to achieve the avoidance of overfitting. The proposed algorithm was tested on Flavia including 32 species and ICL dataset including 220 species, getting 99.37%, 91.2% recognition rates respectively, which were compared with the results of other researchers. We also analyzed the effects of number of train samples, number of hidden layer units, Batchsize etc to the recognition rate of deep belief networks through experiments, providing a basis for good network design.3. Pre-training process has a big influence on the DBNs, so we proposed two modified pre-training algorithms to improve the recognition rate. The first one was called Mean-DBNs, which used the average of parameter training values as the result of pre-training step. This algorithm made the recognition rate to be 93.1% for 220 species of ICL dataset. The second one was defined as PID-DBNs, which added integral element and differential element in the process of parameter adjustment. The experimental results showed that the recognition rate of this improved algorithm can reach to 94.1% for 220 kinds of leaves. Moreover, it completed the training process in a shorter time and had stronger robustness than the original one.