Research On Pepper Object Detection And Pose Estimation Based On Convolutional Neural Network

As an important economic tree species,pepper has a wide planting area and significant economic benefits.At present,pepper picking mainly relies on manual labor,the shortage of labor force and the immature mechanical picking technology have been restricting the development of the pepper industry.Solving the problem of pepper automatic picking can promote the healthy development of pepper industry.This thesis takes pepper as the research object,and conducts in-depth research on the identification and positioning method of pepper based on deep learning,and provides solutions for the realization of automatic pepper picking.The main contents include the following aspects:(1)Considering the limited movement space of the manipulator and the complex obstacles in the scene,the pepper picking system needs to detect the close-range pepper in the current scene.Aiming at the problem that the existing object detection network is easy to confuse the similar-looking near-field and far-field peppers,and it is difficult to accurately detect the nearfield peppers,a pepper detection model based on Multi-Task Learning with Context Reinforcement(MTL-CR)is proposed.First,using the contextual relationship between the close-range pepper and the branches,the branch segmentation task is added to the YOLOv4 target detection task.The multi-task model guides the global feature to express the appearance information of the pepper and the context information of the branches;then,for An attention module is designed for each task to adaptively adjust the global features to avoid mutual interference between tasks;Finally,in experiments on the pepper dataset,the detection accuracy of the MTL-CR model is improved by 12.28%,17.23% and 30.17% compared to YOLOv4,RetinaNet and SSD.Theoretical analysis and experiments show that the MTL-CR model can reduce the false detection of distant peppers and accurately detect close-range peppers.(2)In the process of automatic harvesting of pepper,after completing the object detection of pepper,the posture of pepper should also be estimated.With the posture information of pepper,the manipulator can be more flexibly controlled to achieve precise grasping.Aiming at the problem that the feature distribution of pepper pose data is relatively concentrated,which leads to difficult identification and insufficient manual annotation samples,a pepper pose estimation model based on self-supervised learning with image transformation(SSL-IT)is proposed.First,image transformation is performed on the pepper pose unlabeled dataset to establish pseudolabels between the dataset and image transformation;then,DenseNet is improved by using the SPP structure to solve the limitation of DenseNet on the size of the input image;then,SPPDenseNet is used as the The feature extraction network uses pseudo-labels as supervision information to train the network and obtain pre-training weights;then,transfer the pre-training weights to the pepper pose dataset of artificial labels for fine-tuning to realize the estimation of the pepper pose;finally,Experiments on pepper pose data show that the SPP-DenseNet of the SSL-IT model is improved by 11.3%,7.1% and 1.9% compared to MobileNet-v2,ResNet18 and DenseNet with similar parameters.Based on SPP-DenseNet,the self-supervised learning method of the SSL-IT model improves the accuracy by 16.1% compared with the supervised learning method.Theoretical analysis and experiments show that the SSL-IT model can significantly improve the accuracy of pepper pose estimation by using limited artificial labels and a large number of pseudo-label datasets.