Research On Weeds Identification Based On K-means Feature Learning

Weed identification technology is the key prerequisite for variable spray technology. However current research on it usually based on these handle-crafted features such as color,shape,texture, spectrum which are blindness and depend on human experience, thus may cause differences of weed identification accuracy. This paper utilized K-means feature learning to learn hidden features of the plant automatically, and on this basis contructing weed identification model to achieve this goal of weed automatic identification in the field. The work and conclusion of this thesis are as follows:(1) Image pre-processing. According to the characteristics of the experimental data, this paper adopted two pre-processing steps of image normalization and image whitening. The experimental results show that the data with pre-processing not only have more independence but also eliminating the redundant information and retaining the effective features of raw image. When using clustering algorithm to process the raw data and the data with pre-processing, this paper finds the former hava fuzzy clustering center but the latter have obvious edge features, the data with Principal Component Analysis(PCA) whitening have less distinct features than Zero Component Analysis(ZCA) whitening data; the experiments also prove that after pre-processing, the data have more identification accuracy than raw data.(2) Constructing weed identification model based on K-means feature learning method. Based on the unsupervised feature learning identification model, this paper constructed feature dictionary with K-means feature learning after the pre-processing, then utilized this feature dictionary to extract features from labeled data and train the classifers. The experimental results show that large number of clustering centers, small size of local receptive field and dense feature extraction are critical to achieving high performance. An another layer was added with the single layer, the model can learn higher features and improve the accuracy of identification. Compared with the accuracy of single layer, the second layer’s increased by 1.25% with Support Vector Machine(SVM) multi-class classifier while has 1.88% increase with Multinomial Logistic Regression(Softmax Regression) classifer.(3) Contructing weed identification model based on K-means with convolutional neural network. Taking advantage of the multilayer and fine-tuning of parameters of the convolutional neural network, this paper set K-means unsupervised feature learning as the pre-training, and replaced the random initialization of weight parameters in traditional convolutional neural network. This method make the parameters are set reasonable value before optimization with back propagation. The experimental results show that this method with K-means pre-training achieved accuracy with 92.89%, beyond 1.82% than convolutional neural network with random initialization and 6.01% than the two layer network without fine-tune.In conclusion, in the condition of reasonable parameters, the accuracy may be higher with multilayer and parameters fine-tune, while the method with K-means unsupervised feature pre-trainging can improve identification accuracy further.