Research On Fruit Object Detection Method Of Picking Robot Based On Deep Learning

Realizing the automation and intellectualization of fruit picking is the inevitable choice of modern agriculture,and its key technology is the object detection of fruit by the visual perception system of the picking robot.Accurate and efficient fruit object detection method can improve the picking efficiency of the picking robot,emancipate the agricultural labor force and reduce the economic cost of production.However,there are many complex and uncertain factors in picking objects in natural scenes,which greatly increase the difficulty of fruit object detection.The two difficult problems are small size fruit object detection and occlusion fruit object detection.As a kind of data representation method with strong robustness and migration,deep learning has advantages in feature extraction of complex objects.In order to achieve higher detection accuracy and speed,this thesis studies the two difficult problems mentioned above,and applies the method of deep learning to detect fruit objects.Aiming at the problem of small size fruit object detection,this thesis improves the ability of YOLOv2 algorithm to detect small size target.Firstly,the characteristics of YOLOv2 loss function are analyzed and summarized,and the way to determine the weight coefficient of confidence loss in the target region in the loss function is optimized.A new calculation formula of location prediction loss weight function in the loss function is proposed,which can guide the deep neural network to enhance the sensitivity of small size fruit object in training parameters.Then considering the important role of anchors in target location,the selection rule of anchor size is improved,and a statistical method is designed to determine the number of anchors,so the anchors which are more suitable for small size fruit object are selected.Experiments show that the improved method not only keeps the detection speed of the original algorithm,but also effectively improves the detection accuracy of small size fruit object.Aiming at the problem of occlusion fruit detection,a single stage object detection algorithm named SS-FRCNN is proposed.Firstly,the superiority of Faster R-CNN target detection framework is analyzed.Combining the characteristics of the complex application background of occlusion fruit target detection,a new object detection algorithm SS-FRCNN is proposed and designed based on Faster R-CNN.The design idea and the whole frame structure are described,which make the whole object detection process more compact.Then,the structural details of SS-FRCNN feature extraction network and candidate nomination network are determined.Finally,the training and testing of SS-FRCNN is completed based on the occlusion fruit dataset provided by myself.Experiments show that the proposed SS-FRCNN algorithm is superior to the object detection accuracy of the fruit occlusion,and the object detection speed is improved by 33%.