Retraction notice to “Distilled and filtered deep neural networks for real-time object detection in edge computing” [Neurocomputing 505 (2022) 225–237]

As a branch of cloud computing, edge computing has received extensive attention. Edge computing offloads computing-intensive tasks to the computing-capable devices (i.e., edge servers) at the network edge of the cloud, and the edge servers assist in mobile devices to perform data analysis, which can significantly reduce the execution time of computing-intensive tasks on mobile devices. However, due to the poor quality of wireless channel communication, it may degrade the overall quality of the edge offloading when the mobile devices are linked to edge servers, that is, the communication link between the mobile device and the edge server may become a bottleneck. In this study, we propose a new edge computing scheme to support remote object detection by splitting the deep neural network (DNN), and we design a framework to reduce the amount of data transmitted over wireless links. Specially, the DNN is divided into a head modular and a tail modular, which are deployed on mobile devices and edge servers, respectively. Furthermore, the wireless link transmits the output of the last layer in the head modular to the edge server, instead of sending all the initial tensors of images to the edge server. And we employ three object detection models with complex structures, that is, Faster RCNN (regions with CNN features), Mask RCNN and Keypoint RCNN as the basic models, and modify their network architectures from the following aspects: 1) we introduce the information bottleneck principle and split the network architecture into a head modular and a tail modular, and we perform knowledge distillation to compress the head modular; 2) we embed a neural filter (NF) into the head modular to pre-filter images with no detection objects by using a simple convolutional neural (CNN) network, and the NF performs a binary classification task in the head modular. Lastly, we conduct experiments on the real COCO 2017 dataset and the PASCAL VOC 2012 dataset, and the results show that the proposed method outperforms other baselines. Especially, in person keypoint detection, the mAP (mean average precision) of the proposed method is at least 1.64% higher than other baselines on the COCO 2017 dataset. In addition, the prediction latency of the proposed splitting scheme at the data transmission rate of 1Mbps is at least 49.17% lower than that of local computing scheme.