Abstract:

Camouflage technology can effectively reduce the visual significance of the objects, which conduce a huge challenge to camouflage objects detection. Based on RetinaNet detection framework, spatial attention and channel attention modules are embeded for camouflage features extraction, and a new prediction box filtering algorithm is built based on positioning confidence score, which effectively realizes the detection for camouflage personnel. Experimental results on the extended camouflage dataset show that the average detection accuracy of the proposed model reaches 93.1%, which improves by 8.7 percentage points.

Key words: object detection, camouflage, attention mechanism

摘要：

迷彩伪装技术能有效降低目标的视觉显著度，对迷彩目标检测任务造成巨大的挑战。在RetinaNet检测框架的基础上，针对迷彩目标特性嵌入了空间注意力和通道注意力模块，并基于定位置信得分构建了新的预测框过滤算法，有效实现了对迷彩伪装人员的检测。在扩展后的伪装人员数据集上的实验表明，该模型将检测精度提升了8.7个百分点，达到了93.1%。

关键词: 目标检测, 迷彩伪装, 注意力机制