Research On Visual Recognition And Positioning Of Agricultural Product Based On Multi-channel Neural Network

Agricultural production in our country has gradually shifted from small-scale peasant production to large-scale agricultural planting.Mechanization,automation,and intelligence are necessary conditions for large-scale planting.Agricultural intelligent robots can reduce labor costs in the process of agricultural production and improve the standardization of agricultural production.The research on the visual recognition of agricultural picking robots is an important prerequisite for realizing agricultural intelligent production,which guides the picking robot to complete the picking through providing maturity and location coordinates of agricultural product.This paper takes the daylily picking robot as an example to study the visual recognition and positioning method of the picking robot.The main research contents are as follows:At first,the angle and installation position of the camera on the picking robot is determined according to the characteristics of a large number,small target,and dense growth of day lily,In this paper,the mechanical structure and control system of the picking robot are described and designed briefly to adapt to the characteristics of the data set and target growth which improve the recognition accuracy further.A path recognition algorithm for picking robots is proposed based on color features,and the workflow of picking robots is determined.The software for making data sets are written that label the position information of all target key points in the image.Second,a large number of images of the daylily in the planting area were collected,and the visual recognition results were represented by predicting the coordinates of the target key points.According to the characteristics of the daylily in the image,the mathematical expression of the key points is designed,and all the images collected are marked accordingly.The annotation information includes growth angle,target length,key point position and so on.A multi-channel convolutional neural network model is established to predict the key point position of the daylily target in a bottom-up manner,which aims at solving the problem that traditional target detection cannot judge the growth posture.Based on YOLOv6,the output layer network model of which is replaced,and the angle and length prediction of daylily targets was added to the model.After the feature pyramid is exported by the backbone network,different network branches are responsible for detecting the classification results of the heat map,angle and length,and a correction method was proposed to finally obtain the key point coordinates of the target.Third,the accuracy evaluation indicator of the model is established according to the error between the labeling results and the detection results,and the prediction results of the heat map,angle and length are evaluated by using this indicator.The maximum allowable error is set to determine whether the target is detected,and the detection accuracy of the model is obtained accordingly.Finally,the positioning accuracy of the model is calculated,and the total accuracy of the model is obtained by multiplying it by the detection accuracy.Forth,the key points of the target are located by the binocular camera and its three-dimensional coordinates are calculated.The internal parameters,distortion coefficients and external parameters of the camera are obtained by zhang’s calibration method.SGBM is selected for stereo matching to obtain the disparity map and depth map of the binocular camera.The positioning results were counted and the error of the three-dimensional coordinates was calculated.Finally,the length of the daylily is obtained to judge its maturity by the three-dimensional coordinates of its key points,and the accuracy of the method is verified by experiments.