A Study Of Deep Neural Networks With Application To Plant Leaves Image Recognition

Plants are everywhere in our lives. They are the fundamental part of life on Earth. The first step in botanical study is accurate classification of plants. However, current plant leaf image classification still requires hand-crafted labeling, which is often very time consuming and expensive to obtain, as they require the efforts and expertise of human annotators. Recent advances in deep learning technique have made it possible to automatically extract high-level features from raw sensory data without the help of features handcrafted by domain experts. Therefore, this paper focuses on how convolution neural network on plant leaves image recognition, and builds a full convolution network model. Concrete research content is as follows.Firstly, this paper presents a method for image feature extraction with full convolution neural network. We first pre-process images, including image segmentation, enhancement, rotation and perspective, so as to reduce the amount of computation. We extract features of the images automatically by convolution neural network algorithm. Finally, we use the MLP to replace traditional linear filters. Nonlinear convolution layer of a multilayer perceptron composed of the local data input are mapped by a nonlinear activation function, fully connected network is very easy to over-fitting rather than a linear convolution weights due share itself with some ability to prevent over fitting. Thus we can effectively separable from the nonlinear data.This paper also presents PReLU replacing to Re LU depth instead of the traditional neural network. ReLU(corrected linear function) speeds up the convergence of the training process, which is better than traditional S-shaped in accelerating convergence. But training in plant leaf images will appear in the case of diffusion gradient, so as used herein, activation function PRe LU, it can learn adaptive parameter rectifier, and improve the accuracy and negligible extra computational cost.Through experiments we finally verify the feasibility of the depth of convolutional neural network image recognition in plant leaves of the problem. Comparing with other existing neural network algorithm, we also analyze convolutional neural network performance in real application problems.