Sparse representation for weed seeds classification

In agricultural industry, there is a longing for highly efficient and reliable seeds classification methods. Fast implementation of the existing methods is of great economical importance. Almost all categories of weed seeds have different size, shape and texture, and even the same species are quantitatively diverse in feature. Therefore, feature extraction is a tough, time consuming and labor-intensive task. In this paper, we use the compressive sensing theory, which has been applied to the field of machine learning, to do some dimension reduction treatment such as principle component analysis, downsampling and random projection to avoid careful selection of the feature set. As long as the dimension of the extracted features is beyond the theoretical threshold, we can achieve the desired classification results. It is worth mentioning that on account of humidity, bacteria and many other factors,the seeds are prone to have mould blocks or scabs. Thus, it is extremely necessary to do some simulations like contiguous occlusion. According to the experimental results, we can see that this algorithm is fit for solving this problem.