Computer Vision Based Weed Seeds Recognition

Weeds act as a fatal role in agricultural production, crop growth, invasion of alien species and ecological balance, they bring tremendous negative impact to agricultural and ecological balance. Seed is a significant way of species spreading and weed seeds are the key to mitigating and eliminating the negative effects of weeds. For the purpose of achieving high efficiency and low costs in the task, this paper propose a computer vision based weed seeds recognition method. Geometric normalization is done on weed seeds images via image processing technology, and different kinds of computer vision based feature are extracted, then the feature dimension reduction and feature combination is done, finally the random forest classification model used for robust weed seeds recognition is proposed. The main research issues and results are as follows.(1) The preprocessing of weed seeds images. The original weed seeds images contain a mass of noise and redundancy, and the dimension of the images is too high to do classification directly. Such avoidable noise and redundancy must be removed, for this purpose, we propose a geometric normalization method including image rotation, object detection and image compressing for the preprocessing task. The key point in the task is the image rotation degree determination, a principal component analysis method is engaged in this work. And different ratios of continuous noise are injected into the images to simulate the damaged weed seeds in practical cases.(2) Different kinds of computer vision based feature extraction of weed seeds images. Four kinds of computer vision based features are extracted from the weed seeds images after preprocessing, they are raw pixel values of R channel, G channel and B channel, histogram of oriented gradients, Gist and Sketch Tokens. And feature dimension reduction task is accomplished via kernel principal component analysis. The combination of different kinds of features is also discussed. After feature combination we obtain 15 kinds of features: GistHog, GistHogSTs, GistHogSTsRgb, GistHogRgb, GistSTs, GistSTsRgb, GistRgb, HogSTs, HogSTsRgb, HogRgb, STsRgb. In the experiment the classification performance of different features is discussed. And the experiment results showed that with image damage-ratio equal lesser than 10% the feature extracted from color image and the feature extracted from corresponding gray image provided discriminative power complementarity, the best recognition rate from the combination feature is 3.61 percent higher than single type feature.(3) The robust random forest classification model used for weed seeds classification. The working principle, the parameter selection and its function of random forest are discussed in the paper. In the experiment we verify the effectiveness and robustness of random forest model used for damaged weed seeds recognition, and the support vector machine model is used for comparison. The experiment results showed that random forest classification model handles weed seeds images recognition work effectively, ideal robustness on the damaged weed seeds images is obtained via this model. And the random forest model using linear split function and information gain target function showed better classification performance than other 5 kinds of random forest models.(4) The analysis of time and space performance difference of different storage structure used during decision tree training procedure. The experiment results showed that the random forest model using the sequential compact storage structure during the training procedure had lower memory usage(the used memory decreased 41.34%) and higher time costs(the training time increased 10.25 times) than the one using the sequential storage structure. This high time costing problem may be solved by parallelization technology.