EAWNet: An Edge Attention-Wise Objector for Real-Time Visual Internet of Things

Figure 2

Overall network architecture of EAWNet comprising four parts: (a) backbone architecture, (b) multipath refinement fusion unit, (c) attention modules, and (d) neck consisting of the network. In addition, inside the dotted lines is the explicit implementation of the counterparts. (a) The backbone is responsible for the multiscale feature extraction and is optimized by pass-wise connection to avoid gradient disappearance. (b) The multipath refinement fusion (MRF) unit is responsible for making fusion from the edge prior which is extracted from the ground truth and refined patches. (c) The attention modules learn information of category and structure wisely quickly aided by the edge prior. The position-wise and channel-wise attention modules (in the dotted lines) consist of the attention modules in a parallel manner. (d) The neck is the decoder for object detection which is modified into rotated bounding boxes for better visual effects. All four parts are illustrated in the following sections.