Classification of seeds based on a novel multiple feature combination descriptor

There are a huge number of different species of weeds in agriculture, which have a great negative impact on agricultural development. Therefore, we want to find an efficient, robust algorithm to automate the differentiation of weed seeds. In this paper, we built a data set LZUPSD autonomously, containing 88 species of grass seed images. We processed the images in the dataset in depth and used various feature descriptors on them. We explored how to design a feature combinations of feature descriptors for seed image classification and proposed a new feature LSG based on co m bination of traditional feature descriptors (LBP, SIFT and GIST). The efficacy of LSG was evaluated both in isolation and by combining several of them by means of support vector machine. Based on these comparisons, we further analyzed the effects and relationships of various features on the classification of grass seed images, which has important implications for grass science research. The LSG feature has achieved an outstanding performance in the multi-classification task, laying the technical foundation for smart forestry and grass discipline.