ER-YOLO

In response to the challenges faced by existing object detection algorithms in the application to armored vehicle pose detection, including the inability to effectively utilize edge information and the difficulty in processing multi-scale feature inputs, this paper presents a novel Edge-Refined YOLO (ER-YOLO). To address the issues above, this paper introduces the ER-YOLO framework. First, we designed the Multi-Scale Edge Info Generator (MSEIG) and Conv Edge Fusion (CEF) modules. The MSEIG extracts multi-scale edge feature information from images at shallow layers of the backbone network. Then the CEF module fuses the edge information with deep feature information. Next, we designed the Efficient Parallel Deep Convolution (EPD Conv) to reconstruct the backbone network. This is achieved by decomposing large-kernel depth convolutions into four parallel branches, thereby expanding the receptive field of the network and optimizing the feature extraction capability for multi-scale inputs while maintaining detection performance. Finally, we proposed the Adaptive Region-Edge Awareness Model (AREAM) attention module for the lower layers of the backbone network. This module dynamically filters image regions based on edge and complementary information transmitted from the previous layers, achieving a content-aware sparse attention mechanism.