Multi-target Recognition And Location Based On Deep Learning

To liberate labor and realize the intelligent operation of orchards,fruit and vegetable picking robots gradually become a current research hotspot.I have completed the apple contour and picking point segmentation,matching of apple images,measurement of fruit diameter,localization of picking points and obstacles,experiments on the localization accuracy of picking points and experiments on the effect of light intensity changes on localization accuracy by studying the problem of identification and position of multiple targets such as picking points and obstacles.Aiming at the problem of poor fruit contour segmentation under occlusion,the U-net convolutional neural network method applicable to contour segmentation is used.To solve the problem of redundant pseudo contours due to unclear extraction of important features of images when segmenting apple contours using U-net,the attention mechanism sc SE module is introduced into U-net to improve the efficient extraction and utilization of important features.In addition,a weighted multilevel cross-entropy loss is used as the loss function of this improved network to measure the degree of mismatch between the predicted and true values of the training model,and the training results are compared with those of U-net to verify the effectiveness of the improved network for apple contour segmentation under occlusion,and the apple contour segmentation of fruits under occlusion is finished.To solve the problem of low accuracy and recognition success rate of fruit picking point segmentation,Seg Net convolutional neural network method applicable to point and block segmentation was used.Since it appeared in the experiment that the picking point segmentation was unclear due to feature loss occurred when the input information was passed between multiple layers,the idea of Dense Net was introduced into Seg Net to realize feature reuse several times,and the training results of Dense-seg Net and Seg Net training results are experimentally compared to verify the effectiveness of the improved network for picking point segmentation and improve the accuracy of picking point segmentation and recognition success rate.Since the current stereo matching algorithm cannot guarantee the efficient and high-quality output of parallax map and the mismatching of homonymous points in binocular images due to illumination changes,a matching algorithm incorporating optimized Census transform is improved based on SGM algorithm,which can effectively reduce the effect of illumination changes on the matching between pixel points,and the improved algorithm is verified by comparing with the comparison BM and SGBM algorithms.The effectiveness of the improved algorithm is verified by comparing with the comparison BM and SGBM algorithms.The problems of localization of picking points and obstacles,the accuracy of localization,and the verification of whether the Census transform is resistant to the effects of lighting changes are investigated.In this paper,the improved algorithm based on SGM is used to complete the 3D reconstruction of the apple tree model and realize the positioning of obstacles;the images containing only fruits and picking points are extracted by contour segmentation and picking point segmentation,and then the 3D reconstruction is performed to realize the measurement of fruit diameter and the positioning of picking points,and the point cloud coordinates are saved in file format,and the positioning accuracy of the improved stereo matching algorithm is verified through experiments It is experimentally verified that the improved SGM algorithm meets the picking requirements and that the algorithm has the ability to resist the influence of light changes.