基于注意力和上下文的多尺度图像背景下的小目标检测方法

Small Object Detection Method under the Background of Multi-Scale Image Based on Attention and Context

在多尺度多目标的背景下,小目标由于像素少、提取特征困难,其检测精度远远低于大中目标。文章通过使用离散自注意力提取跨尺度的全局的上下文背景信息,使用跨尺度通道注意力和尺度注意力来增强模型的尺度敏感性,捕捉到更多不同的、更丰富的物体-物体、背景-物体信息,使得每一层特征层都是一个跨空间和跨尺度的拥有更丰富特征信息的特征层,从而提高在多尺度背景下小目标检测的效果。在COCO数据集上,本算法的APs高于基准retinanet最高达2.9,在DIOR数据集上mAP能够达到69.0,优于该数据集上最优算法,同时能够维持自己单阶段的速度。

Under the background of multi-scale and multi-target, the detection accuracy of small targets is far lower than that of large and medium targets due to fewer pixels and difficulty in feature extraction. Through using discrete self-attention to extract cross-scale global context information, and using cross-scale channel attention and scale attention to enhance the scale sensitivity of the model, this paper captures more different and richer object-Object and background-object information, so that each feature layer is a feature layer with richer feature information across space and scale, thereby improving the effect of small target detection under the background of multi-scale. On the COCO data set, the APs of this algorithm are higher than the benchmark retinanet by up to 2.9, and the mAP on the DIOR data set can reach 69.0, which is better than the optimal algorithm on this data set, while maintaining its own single-stage speed.