基于锚框的高效单层特征目标检测  

Efficient Single Layer Feature Target Detection Based on Anchor Box

作　　者：崔小同 王博[1] 李安琪 CUI Xiaotong;WANG Bo;LI Anqi(Liaoning Petrochemical University,Fushun 113001,China)

机构地区：[1]辽宁石油化工大学,辽宁抚顺113001

出　　处：《现代信息科技》2023年第22期40-43,47,共5页Modern Information Technology

摘　　要：目标检测广泛应用在公共场合的智能监控、自动驾驶与计算机辅助诊断等领域。文章提出了单层特征目标检测替代复杂的特征金字塔结构,从而提升模型的推理速度和预测精度。在模型搭建过程中,瓶颈特征结构采用了单层空洞残差编码器,样本选择采用了统一匹配机制,并采用了任务对齐检测器。在COCO(Microsoft Common Objects in Context)数据集下,大量实验证明该方法的有效性,以Res Net50为基准,预测精度达到了38.2 m AP,比Retina Net的推理速度快1.4倍,精度提高2.3 m AP。该模型具有推理速度快、预测精度高等特点,可以应用在许多特定场景中。Object detection is widely used in fields such as intelligent monitoring,autonomous driving,and computer-aided diagnosis in public places.This paper proposes a single-layer feature object detection method to replace the complex feature pyramid structure,in order to improve the inference speed and prediction accuracy of the model.During the model building process,a single-layer cavity residual encoder is used for the bottleneck feature structure,unified matching mechanism is used for the sample selection,and a task alignment detector is used.Under the COCO(Microsoft Common Objects in Context)dataset,a large number of experiments have demonstrated the effectiveness of this method.Based on ResNet50,the prediction accuracy reaches 38.0 mAP,which is 1.4 times faster than RetinaNet's inference speed and improves the accuracy by 2.3 mAP.This model has the characteristics of fast inference speed and high prediction accuracy,and can be applied in many specific scenarios.

关 键 词：目标检测 单层特征 特征金字塔 编码器 任务对齐 

分 类 号：TP391.4[自动化与计算机技术—计算机应用技术]