Research On Identification Method Of Weed Seeds Based On Interlacing PCANet

The identification of weed seeds is very important to improve resistance of crops to pests and to increase the yield of crops. Existing seed identification methods based on machine vision extract features from images based on artificial assumption and it’s difficult to extract new features because of professional knowledge in the related field are required. In the view of machine learning, the discriminative features of images should be learned from image samples, rather than artificially defined. Features of deep learning is learned from image by machine automatically. Deep learning model can extract higher level features from the input training image samples layer by layer through simulating human brain’s working mechanism. These features have better representation ability to image and it have been exhibited superior generalization ability on many datasets. So the deep learning network is suitable for the weed seed images identification which has large within-class difference and can improve the accuracy of weed seed images identification methods based on machine vision. Based on the above analysis, this paper construct a deep learning network for weed seeds identification problem which can automatically extract higher level features from seed images and use these features represent images.The main research contents and results of this paper are as follows:(1) PCANet is a simple convolutional neural network and this network initial weight through principal component analysis method. Compared with the convolutional neural network which initial weights through random initialization method, it reduce the training parameters and accelerate the training speed. Interlacing PCANet improve the connection between input and output of PCANet from one-to-many to many-to-many. Interlacing PCANet can extract more discriminative features from images in higher layer through the combination of lower level features and reduce over fitting through the sparse connection of network. The experiment results show that, compared with PCANet, the recognition rate of interlacing PCANet on weed seed images is improved by 2%.(2) This paper uses sparse orthonormal transforms method to train the weight through reduces the reconstruction error of the matrix composed by the image patches. Compared with deconvolution method and Back Propagation(BP) algorithm, this method accelerates the training speed of weights. Through the reconstruction of patches matrix composed by the patches in the first layer of interlacing PCANet can increase the relation of the output in each layer with training samples and reduce the impact of the information loss of the lower layers to the higher layers. The experiment results show that, compared with the network without training the weight, the recognition rate of network with weight training process on weed seed images is improved by 4%.