Research On Soybean Weed Recognition Based On Image Processing

Soybean is an important oil crop,high protein food and feed crop in China.Its planting area is second only to rice,corn and wheat.In recent years,due to the influence of climate change,diseases,insect pests and grass damage,the soybean planting area in China has decreased year by year,and the output has decreased year by year.Chemical weeding is widely used in the field of agricultural production due to its labor-saving,labor-saving,easy-to-operate,and good weeding effects.When spraying herbicides with drones,uncontrolled spraying on large areas will not only cause waste,but also endanger the safety of agricultural ecosystem.In order to solve the problems of noise interference and serious image information loss during the UAV image acquisition process,which leads to low accuracy of crop weed identification,this paper takes soybean seedlings,gramineous weeds and broad-leaved weeds as the research objects.The method of combining image processing and BP neural network is used to identify the soybean weed images,in order to improve the accuracy of crop weed identification,and provide a technical reference for drone precision spraying.The main research contents of this article are as follows:(1)Research on image denoising,enhancement and segmentation algorithms.Aiming at the problems of unclear jitter and noise caused by equipment transmission when the UAV is collecting images,compared with the traditional single image denoising algorithm,it is found that the mixed filter denoising method is better than the single image denoising method,and the extreme value adaptive median filtering and wavelet filtering are combined to denoise the image.By comparing and analyzing the enhancement effect of histogram equalization enhancement and multi-scale Retinex enhancement algorithm on soybean weed images,the Retinex algorithm was finally selected for image enhancement processing.In order to better extract the leaf feature information,the Otsu algorithm is used to select the binary threshold to segment the image background,and the morphological image processing method is used to further segment the slightly overlapping leaves.(2)The method of image contour,shape and texture feature extraction is studied.The Canny edge detection algorithm is improved.The traditional Canny algorithm needs to set filter parameters,high and low thresholds,and poor detection results in salt and pepper noise environments.The combination of adaptive median filtering and Gaussian filtering to replace the traditional Canny algorithm used for Gaussian filter denoising,adding gradient calculation template and introducing dichotomy in Otsu algorithm to find high and low thresholds to improve the Canny algorithm.Sobel operator and Laplacian operator remove a large number of false edge points,and the detection effect is better.The contour information is used to further extract the shape features of the blade,and the width-to-length ratio,circularity,rectangularity,and invariant moment are selected as the parameters describing the shape of the blade.The leaf texture features are extracted through the gray level co-occurrence matrix,and the entropy,inverse moment,energy,contrast and correlation are selected as the characteristic parameters describing the leaf texture.(3)Constructed soybean weed recognition model based on BP neural network.First of all,using shape features,texture features and mixed shapes and texture features as input parameters,three weed recognition models based on BP neural network are constructed.Experiments show that the combination of shape and texture features is more accurate than using any one of the features alone for recognition.The rate is higher.Secondly,using the shape and texture features as input parameters,comparing the recognition effects of the traditional recognition model and the model constructed in this paper,the accuracy rate of the traditional weed crop recognition model is 91.11%,and the accuracy rate of the model recognition in this paper is 96.67%,an increase of 5.56%.