基于BoT-YOLOX的毫米波图像目标检测
Object Detection in Millimeter Wave Images Based on BoT-YOLOX
-
摘要: 主动毫米波(active millimeter wave,AMMW)图像具有噪声多、易含伪影、小目标多等特点,一直是隐匿目标检测的挑战.为此,提出了一种基于BoT-YOLOX的毫米波图像目标检测方法.首先,在模型主干网络中引入瓶颈型Transformer (bottleneck Transformer,BoT),加强模型的特征提取能力;然后,调整多尺度目标检测层,并集成全局注意力机制来提高对小目标的检测能力;最后,提出一种多视角加权框融合的后处理方法,用于集成不同视角检测结果,以提高模型的鲁棒性.在自行采集的包括54 000幅图像的AMMW数据集上,与基准模型(YOLOX)相比,该模型达到了93.22%的检出率和4.46%的误检率,AP提升了6.74个百分点;在公开AMMW数据集上,与主流方法相比,mAP提升了4.07个百分点.实验结果表明,所提方法对AMMW图像场景的目标,小目标检测准确度更加出色.Abstract: Active millimeter wave (AMMW) images are characterized by high noise, artifacts, and small objects, which has always been challenges for concealed object detection. Therefore, a method is proposed for detecting objects in millimeter-wave images based on BoT-YOLOX. Firstly, Bottleneck Transformer (BoT) is introduced into the model backbone network to enhance feature extraction capability of the model. Then, multi-scale object detection layer are adjusted, and global attention mechanism is integrated to improve detection ability of small objects. Finally, a post-processing method of multi-view weighted boxes fusion is proposed to integrate the detection results of different views to improve the robustness of the model. On the self-collected AMMW dataset, which includes 54 000 images, compared with the baseline model (YOLOX), the model achieves a detection rate of 93.22% and a false detection rate of 4.46%, and AP is increased by 6.74 percentage points. On the public AMMW dataset, compared with mainstream methods, the mAP is increased by 4.07 percentage points. The experimental results show that the proposed method is more accurate in detecting small targets in AMMW image scenes.