激光与光电子学进展, 2020, 57 (6): 061008, 网络出版: 2020-03-06
一种双注意力模型引导的目标检测算法 下载: 1224次
Object Detection Algorithm Guided by Dual Attention Models
图像处理 目标检测 卷积神经网络 小目标 注意力模型 image processing object detection convolutional neural network small objects attention model
摘要
为了解决对小目标物体识别精度较差的问题,提出了一种双注意力模型引导的目标检测算法。该方法基于单阶段检测算法的实现原理,通过引入两种注意力模型来提升检测性能,尤其是对小目标物体。首先在卷积神经网络中引入一个多尺度特征级联注意力模块,对原始卷积神经网络的特征图中各区域进行不同重要程度的关注,降低特征图的背景及负样本信息的干扰,特别是在浅层特征图中可对小目标物体进行有效的关注。此外,密集连接的方式缓解了网络反向传播过程中梯度消失的问题。其次,对融合后的特征引入显著通道自注意力模块,区分特征图不同通道,筛选出有用的通道信息,使待检测的特征更具表征性。在目标检测基准数据集COCO上进行测试,验证了所提方法的有效性和先进性。
Abstract
In order to solve the problem of inferior recognition accuracy for small objects, a object detection algorithm guided by dual attention models is proposed. The method is based on the realization principle of single-stage detection algorithms, and introduces two attention models to improve the detection performance, especially for small objects. Specifically, a multi-scale feature cascade attention module is first introduced into the convolutional neural network, which weights the importance on different regions of the original convolutional neural network's feature map to reduce the interference of background and negative object information in the feature map, especially highlighting the small objects effectively in the shallow feature map. Besides, dense connection alleviates the problem of gradient disappearance in the process of back propagation. A salient channel self-attention module is introduced for the fused features, which focuses on the difference among different channels of the feature map so as to screen out useful channel information, thus making the feature map to be detected more representative. In addition, the experiments on COCO benchmark dataset of object detection verify the effectiveness and advancement of the proposed method.
冀中, 孔乾坤, 王建. 一种双注意力模型引导的目标检测算法[J]. 激光与光电子学进展, 2020, 57(6): 061008. Zhong Ji, Qiankun Kong, Jian Wang. Object Detection Algorithm Guided by Dual Attention Models[J]. Laser & Optoelectronics Progress, 2020, 57(6): 061008.