Abstract: In order to solve the application problems facing traditional industrial robots such as difficulty in identifying complex industrial parts and singleness in recognition, a visual recognition and sorting strategy based on transfer learning is proposed. Firstly, the pictures taken by high-precision industrial camera are processed by Haclon software, such as expansion, corrosion and then imported to Pytorch which has prepared a good neural network model to identify the target classification, and finally the purpose of industrial robot sorting is achieved. In the experiment, two kinds of mushrooms with varied shapes are taken as an example to implement the transfer learning and sorting on the UR5 robot platform. The experimental results illustrate the good accuracy and stability of the proposed algorithm.

Key words: transfer learning, vision recognition, image processing, neural networks, sorting

摘要： 针对传统工业机器人辨识复杂工件困难、识别度单一等问题，提出一种基于迁移学习的视觉识别与分拣策略。高精度工业相机拍摄到的图片经过HALCON软件图像膨胀、腐蚀等处理之后，导入Pytorch中的神经网络模型，利用迁移学习对目标进行识别分类，最终实现工业机器人智能分拣的目的。实验中，在UR5机器人平台上以形状多变的两种菇类为对象进行迁移学习，进而完成识别及分拣。实验结果表明该策略具备良好的准确性和稳定性。

关键词: 迁移学习, 视觉识别, 图像处理, 神经网络, 分类