双支路注意力特征融合的卷积稀疏编码目标检测

Convolutional sparse encoding target detection based on double branch attention feature fusion

针对现有目标检测模型在实际运用中会受到各种噪声的影响而导致性能退化的问题,提出一种双支路注意力特征融合(double branch attention feature fusion,DBAFF)的方法。基于CenterNet的结构设计,引入卷积稀疏编码(convolutional sparse coding,CSC)去噪模块。通过双支路互补学习,自适应选择不同模态的有效信息,使融合特征达到最优化,有效解决该类模型的退化问题。实验结果表明,该方法在噪声数据集VOC-Nosiy上mAP50、mAP75、mAP性能分别达到了57.9%、29.8%、24.5%,检测速度FPS达到111帧,综合性能优于原网络和仅添加卷积稀疏编码的去噪网络。

To address the problem of performance degradation caused by various noise in practical applications of existing target detection models,a method of double branch attention feature fusion(DBAFF)was proposed.Based on the structural design of CenterNet,a convolutional sparse coding(CSC)denoising module was introduced.Through complementary learning from two branches,the effective information of different modalities was adaptively selected to optimize the fusion feature and effectively solve the degradation problem of this type of model.Experimental results show that the performance of mAP50,mAP75,and mAP on the noisy dataset VOC-Nosiy reaches 57.9%,29.8%,and 24.5%,respectively,with a detection speed of 111 frames per second.The overall performance of the proposed method is superior to that of both the original network and the denoising network with convolutional sparse coding.