(Retracted) Target recognition and localization based on lightweight single-shot multibox detector network for robotics

Abstract

Target recognition and localization are essential in computer vision and pattern recognition in robotics. The artificial extraction of features was omitted with the emergence of deep convolutional neural networks, reducing the influence of human factors on the results. The single-shot multibox detector (SSD) network has achieved excellent recognition by high precision and fast speed in target recognition and positioning. However, some small real-time systems are difficult to implement because of their demanding hardware and extended training time. Based on the digital normalization and residual network structure, the depth-wise separable convolution is proposed to replace the traditional convolution. The improved SSD network structure was used for identification and positioning in our work. The speed of training and testing increased without a decline in the accuracy, thus reducing the dependence on hardware. The method has achieved good results on the PASCAL VOC dataset after testing. It can also be applied to the field of intelligent inspection robots and intelligent security robots.