Research On Robot Grabbing Object Detection Method Based On Lightweight Convolutional Neural Network

Recently,deep learning has achieved great success in computer vision applications,and has been sought after by scholars and Internet companies researchers.As an important basic algorithm,standard Convolutional Neural Networks have been widely used in the field of image recognition and object detection,and various excellent convolutional network structures have been proposed.With the rapid development of society,more and more applications are in urgent need of the introduction of artificial intelligence technology.Deep learning technology has achieved great success,it faces many challenges in many application situation,such as having amount of trainable parameters,heavy hardware requirements,and difficult to guarantee requirements with real-time.In addition,it is easy to overfit and it is difficult to ensure generalization ability.Therefore,the lightweight deep neural network has been developed rapidly.The lightweight methods are to perform deredundant processing on a single model and design a single lightweight model.This article focuses on characteristics of convolutional neural networks,and researches image recognition and object detection of lightweight convolutional neural networks with ensemble learning,so that the model can achieve better generalization performance.The main research works are as follows:(1)Based on the probability vector outputs characteristics of multiple classifier systems convolutional neural network model,the ensemble diversity measure method called Dpv is proposed.The method calculate diversity of classifiers,which can be used in the classifiers the output probability vector,introducing the difference between two vectors.By training multiple convolutional neural networks models with different structures on CIFAR-10 and CIFAR-100 datasets,experimental results demonstrate that the proposed method is similar to double-fault measure,better than the Q statistics and the disagreement measure on CIFAR-10,and this method is the best compared to the other three methods on CIFAR-100.(2)A lightweight convolutional neural network ensemble model called M-Res Net is proposed.First,basic models are obtained when trained and tested on 10-monkeys dataset and 5-flowers dataset,and calculate the diversity of classifiers by means of the DPV method.The experimental results show that there have best diversity between model called Res Net-50 and model called Mobile Net-V2,this obtained ensemble model called MRes Net.Finally,in order to further verify the effectiveness of ensemble model called MRes Net,a experiment will be carried out on five small parts datasets which collected in the simulated industrial scene of mobile robot.(3)This paper carry out the research of lightweight convolutional neural network ensemble systems in the field of object detection about robot parts.This experiments obtained the model of small parts based on the SSD framework,which the ensemble model called M-Res Net is used on the backbone network of SSD,taking different types of screws and nuts as detection objects,using multi-scale training,network pre-training and data enhancement.Results show that the improved SSD framework can achieve a better recognition performance.