Multiple Convolutional Neural Network for Feature Extraction

Recent theoretical studies indicate that Deep Neural Network has been applied to many image processing tasks. However, learning in deep architectures is still difficult. One of the neural network, Convolutional Neural Network (CNN) has gained great success in image recognition and it builds features by automatic learning. More importantly, CNN can operate directly on the gray image, so it can be directly used for processing classification of the image. In order to utilize CNN to recognize plant leaf, a hierarchical model based on CNN is proposed in this paper. We firstly do some pre-processing, such as illumination changes, rotation and leaf distortion. After that, we applied the method of CNN to extract the features of leaves pictures. One focus on our network is about the depth of CNN, which affects the ability of capability of convolution. Thus, we try our best to choose the best depth of CNN with several experiments. Moreover, in order to destroy the symmetry of networks, the strategies used in this paper is to add a mathematical formula for feature map connection between convolutional layer and sampling layer. The experimental results show that the proposed method is quite effective and feasible. And we also applied other classification methods to the ICL dataset. By contrast, our classification is much better than other methods.