果蔬机器人采摘中基于改进SDD模型的目标自动化识别研究

Research on Automatic Target Recognition Based on Improved SSD Model in Fruit and Vegetable Robot Picking

为了提高果蔬机器人采摘的识别准确率,利用SDD网络进行实时检测。对基础网络的构造和检测层默认框进行改进,增加卷积层数量,减少计算参数,使其利于对不同尺度的果蔬进行识别,得到基于改进SDD模型的目标自动化识别框架。分析激活函数,发现Relu函数具有更快的收敛速度,利用Relu函数进行激活操作,并对训练结果进行分析,结果表明,改进后的SDD模型具有更快的收敛速度;当置信度为0.45时,改进后的SDD模型具有较理想的准确率、召回率、FI值和IOU值,可以满足果蔬机器人采摘的识别要求。

In order to improve the recognition accuracy of fruit and vegetable robot picking,SDD network is used for real-time detection.The structure of the basic network and the default box of the detection layer are improved,the number of convolution layers is increased,and the calculation parameters are reduced,so that it is convenient to recognize fruits and vegetables of different scales,and an automatic target recognition framework based on the improved SDD model is obtained.The activation function is analyzed,and it is found that the Relu function has faster convergence speed.The activation operation is carried out by using the Relu function,and the training results are analyzed.The results show that the improved SDD model has faster convergence speed.When the confidence is 0.45,the improved SDD model has ideal accuracy,recall,FI value and IOU value,which can meet the recognition requirements of fruit and vegetable robot picking.