匹配多尺度特征与预测任务的实时目标检测

Matching Multi-Scale Features and Prediction Tasks for Real-Time Object Detection

在基于卷积神经网络的目标检测算法中,浅层高分辨率特征包含更多细节信息,有助于抽象特征完成精确的定位任务;深层特征包含抽象的语义信息,更适合目标存在性预测任务。研究发现,现有的不基于先验框的检测方法直接在同一特征图上预测所有任务时,并没有匹配上述特征与预测任务,而这一特征与任务不匹配的问题限制了检测精度。为解决这一问题,提出了一种匹配目标多尺度特征与预测任务的实时目标检测算法,简称MFT检测器。以CenterNet检测器为基础,同时完成浅层细节特征与精确定位任务的匹配,多尺度多感受野抽象特征与目标存在性预测任务的匹配。实验结果表明,所设计的MFT检测器缓解了特征与预测任务不匹配的问题,从而显著提高了检测精度,且检测速度保持在94.5frame/s,能够保证检测实时性。

In object detection algorithms based on convolutional neural networks,high-resolution features from lower levels contain more detailed information,which can help the abstract features complete the accurate positioning task;deep-level features contain abstract semantic information,which is more suitable for target existence prediction task.When the most existing anchor-free detection method directly predicts all tasks on the same feature map,it does not match the above features and prediction tasks,which limits the detection accuracy.To this end,the MFT detector,a real-time object detection algorithm,is proposed to match multi-scale features and prediction tasks of targets.MFT detector is based on CenterNet detector,which can match shallow detail features with accurate positioning task,and match multi-scale,multi receptive field abstract features with target existence prediction task.Experimental results show that the proposed MFT detector alleviates the mismatch between features and prediction tasks,and significantly improves the detection precision while maintaining a high speed of94.5 frame/s,which meets the requirement of a real-time vision system.