Research On Cascade Network Of Robot Lightweight Target Classification And Grasping Control Based On Deep Learning

With the advent of the global information age,the production output of our industry has reached a high level.The prevalent of AI has promoted the advancement of a large number of industries.The combination of advanced technologies such as deep learning and image processing with robotic has achieved remarkable results.However,there are difficulties such as poor location recognition,and poor real-time performance.Robot classification and grasping detection are of great significance to improve the progress of work production and promote economic development.Aiming at the above problems,a cascade network of target classification and grasping control based on lightweight network is designed.This network can output the category and grasping attitude of the grasping target.Finally,a simulation manipulator platform is built to verify the algorithm.(1)The lightweight network YOLOv4-Tiny is selected as the basic network of the first level target classification and improved.The improved Receptive Field Block(RFB)is added to enhance the receptive field of the network to the detected target;The Convolutional Block Attention Module(CBAM)is added to improve the degree of attention to small target objects;Using Path Aggregation Network(PANet)in the neck network,the structure can combine the deep feature map extracted by the feature extraction network with the shallow feature map to improve the utilization rate of information in the shallow feature map.On the expanded multitarget sample Cornell dataset,the Average Precision(AP)reaches 94.47%,and the detection speed reaches 0.0158s per sample.(2)For the task of grasping attitude detection,Lightweight network Generative Grasping Convolutional Neural Network(GG-CNN)is selected as the base network and improved,which deepens the layers of the backbone network of GG-CNN,and replaces transposed convolution with dilated convolution,so that the network can get the features that are conducive to grasping detection;The Atrous Spatial Pyramid Pooling(ASPP)is introduced,and the improved network has a wider range of receptive fields and can obtain feature information of various scales.Merge and sum the shallow feature maps with the deep feature maps,the network can obtain more semantic and detailed information of the image.The improved algorithm achieves 81.27%AP and 0.0104s detection speed per sample on the Cornell dataset using the new pixel level annotation method.(3)To enhance the properties of the classification network,the Cornell dataset is expanded with multi-target samples to make up for the lack of multi-target scenarios in the Cornell dataset,and the dataset is expanded with data enhancement.GG-CNN series algorithms transform Cornell datasets to adapt to the network,but there are still some problems,such as discrete tags and incomplete capture possibilities.In this paper,a pixel level annotation method is adopted,and the annotation file of the grasping posture of the pixel points is generated for network training.The network is tested on the simulation manipulator,the average success rate of scraping reached 88.72%,which proved that the designed network achieved the target scraping effect well.