Edge-Cloud Collaboration Architecture for Efficient Web-Based Cognitive Services

Many web services are often latency-sensitive, have high network capabilities, and have computing resource limitations, as demonstrated by deep neural network (DNN) model-based web applications. In this work, a DNN model-based web application providing visual cognitive services is implemented and explored to address latency-sensitive problems, taking into account the broad practicability and importance of web object recognition tasks in intelligent applications and modern systems. We propose a collaborative architecture of end-user, edge server, and cloud server, employing a binary offloading strategy to reduce the upload rate of images while ensuring good detection performance, thereby reducing the response time. Detection scenario-oriented data augmentation is carried out on the dataset to improve detection accuracy. Finally, we compare the performance of the proposed approach with traditional object recognition services running entirely on the cloud. The experimental results show that the detection scenario-oriented strategy significantly improves the detection accuracy of fine-trained YOLOv5 models on the PASCAL VOC 2012 dataset. Compared with the traditional cloud-based architecture, the proposed edge-cloud collaboration architecture can detect more objects and reduce response time efficiently.