Abstract:
Abstract: It is of great significance to explore a weed identification method for overlapping leaves to realize precision pesticide spraying, and improve crop yield and reduce environmental pollution. In view of the problems of low recognition rate and invalid shape features for overlapping leaves, a method based on image sub-block and reconstruction was proposed to realize weed identification for spinach. In this paper, the experimental images were taken from the agricultural field of Shanxi agricultural university in October to November 2018, and the growth cycle of the spinach was 21-36 d. A total of 55 images with overlapping areas of spinach and weed were collected, and the image resolution was 4 608×3 456. Firstly, since the experimental images included soil and green vegetation, with the great color differences between the two, the green vegetation should be extracted from the soil background. Two ultra-green models of and were selected to achieve RGB (red, green, blue) images gray-scale of spinach, which were both effective methods to distinguish green vegetation from the background. Then, the iterative threshold segmentation was selected to obtain the green vegetation foreground image, and the holes were filled with the morphology method. Aiming at the failure of shape features for overlapping leaves, fully used the local information advantages of crops and weeds and significantly improved the accuracy of weed recognition, this paper segmented the image into different sizes of sub-blocks, and extracted the features from each sub-block separately, to achieve the purpose of weed recognition. In order to enhance the robustness of image blocks, the divided blocks should be of the same size and have overlapping regions, and then we explored the optimal sub-block size acquisition method to balance the contradiction between recognition time and accuracy. After segmentation, each sub-block image was labeled manually, and the labels had three categories: the black background was labeled as 0, the spinach was labeled as 1, and the weed leaf was labeled as 2. Considering the differences of texture and color features between crops and weeds, 9-dimensional color features of H (hue), S (saturation) and V (value) components based on HSV (hue, saturation, value) color space was obtained; 59-dimensional texture features based on local binary patterns (LBP) algorithm was extracted from each sub-block; 10-dimensional vein features based on fractal box dimensions method was extracted. Thus, a total of 78-dimensional feature vectors was obtained, which can provide favorable experimental data for training and testing the recognition model. After screening, 15 images with prominent overlapping leaves were selected, and each image was expanded into 225 image blocks, so the feature vectors were 3 375×78, then the data used for recognition were 2 026×78 after removing the background image blocks. Finally, a support vector machine (SVM) classifier was constructed for classification and recognition. Cross-validation (CV) was used to evaluate the performance of the classifier, namely the sub-image blocks generated by one image were used as verification sets, and the remaining sub-image blocks were used as training sets. In order to realize the visualization of image sub-blocks, this paper reconstructed the results by image sub-block edge expansion method and voting window mechanism, since the overlap rate of the original image sub-block is 50%, the original N×N image sub-blocks with overlapping regions can be constructed into the (N+1)×(N+1) image without overlapping regions. The experimental results show that the average recognition rate of the proposed method was 83.78%, which is higher than K-Nearest neighbor (KNN), decision tree, naive Bayesian model and ensemble learning algorithm of Adaboost. This study can realize weed identification with overlapping leaves, thus providing a theoretical basis for the development of intelligent lawnmower.