Research on robot target recognition based on deep learning

Abstract

For the traditional machine vision recognition technology in the industrial field can not handle the problem of different classes of workpieces placed randomly and stacked on each other, this paper improves the SSD algorithm model based on the research of deep learning target detection algorithm. Firstly, a depth-separable convolutional structure is introduced to optimize the VGG backbone feature network. Then a multi-level feature fusion mechanism is introduced in the prediction part to increase the semantic information of features. Qualitative and quantitative experimental results show that the improved optimization method of the SSD model in this paper is validated well on the dataset, and the improved SSD model mAP value is increased by 4.3% compared with the original, and the detection speed is increased by nearly two times, thus proving the effectiveness of the improved method.