Weed seeds classification based on PCANet deep learning baseline

There are a large number of various kinds of weeds in agriculture. Weeds have a great impact on the development of agricultural production and agricultural economy. The reproduction and spread of weeds are mainly dependent on weed seeds. So we want to find an efficient algorithm with robust and accurate classification of weed seeds, which has an important practical value and economic significance. PCA Network has been applied to image feature extraction and achieved fantastic effects. Here we propose a variant of PCA Network and apply it to the classification of weed seeds in agriculture. The difference between the proposed method and the original PCA Network lies in that we get L1 families of orthogonal filters rather than one family of orthogonal filters in the second stage of PCA. After using the PCA Network variant method to extract image features, we conduct an experiment to test the classification accuracy of weed seeds. The data sets contain 91 types of weed seeds. In the experiment we use a large margin classifier to construct a linear classifier, which is based on affine hulls. Next we use the features extracted from the test samples to examine the recognition accuracy rate. Experiment results show that the PCA Network variant method obtains good classification results and improves the recognition accuracy. In the data sets composed of 91 types of weed seeds, 45 arrives at 100% recognition rate of classification, and 90.96% average recognition rate. At the same time, our algorithm shows relatively higher robustness than the recognition of weed seeds images does. This algorithm improves the classification accuracy rate of weed seeds greatly and thus can be applied to the agricultural production practice.