基于EfficientNet的无锚框目标检测模型
An Anchor-free Object Detection Model Based on EfficientNet
卜子渝 1杨哲 2刘纯平3
作者信息
- 1. 苏州大学 计算机科学与技术学院,江苏 苏州 215006
- 2. 苏州大学 计算机科学与技术学院,江苏 苏州 215006;江苏省计算机信息处理技术重点实验室,江苏 苏州 215006;江苏省大数据智能工程实验室,江苏 苏州 215006
- 3. 江苏省计算机信息处理技术重点实验室,江苏 苏州 215006;江苏省大数据智能工程实验室,江苏 苏州 215006
- 折叠
摘要
目标检测是计算机视觉的热门研究方向之一,包含分类和定位两个任务.针对单阶段目标检测模型普遍存在的两个问题:训练时正负样本的不均衡以及锚框的设置需要人工干预,提出一种基于EfficientNet的无锚框目标检测模型(Anchor-free Efficientnet-based Object Detector,AEOD).AEOD先筛选出落在目标框中的特征点,再根据特征点所作的预测计算代价矩阵,在训练时基于代价矩阵为目标动态分配正负样本,从而达到平衡二者数量的目的.此模型通过特征图中的特征点直接预测目标的位置和形状,不仅省去了人工设置锚框的环节,还提高了可检出目标的数量.此外,可缩放的EfficientNet进一步提高了模型的泛化能力,使之可以接收多尺度的输入.在PASCAL VOC07+12 数据集中,AEOD最高可以获得91.3%的平均精度(mAP),检测速度达到32.1 FPS,较其他主流的目标检测模型有显著提升.
Abstract
Object detection is one of the hot research areas in computer vision,which includes two tasks:classification and location.Due to the two common problems appearing in one-stage object detector:extreme imbalance between positive/negative samples during training and anchors pre-defined deeply depending on manual settings,an anchor-free efficientnet-based object detector(AEOD)is pro-posed.AEOD first selects out the feature points that fall in the target box,then calculates the cost matrix based on values predicted by these feature points,finally assigns the positive/negative samples to the target dynamically according to the cost matrix during the training.Therefore,the number of positive/negative samples is balanced to enhance the performance of the model.AEOD directly predicts location and shape of the object through the feature points in the feature maps.As a result,not only the step of pre-defining anchors can be skipped,but also the number of objects that successfully detected increases.In addition,the scalable backbone(EfficientNet)improves the generalization ability of AEOD,it can receive multi-scale input.AEOD achieves the highest 91.3%mAP on PASCAL VOC07+12 at speed of 32.1 FPS,showing a significant improvement compared to other modern models.
关键词
深度学习/计算机视觉/目标检测/正负样本分配算法/无锚框Key words
deep learning/computer vision/object detection/positive/negative samples assignment algorithm/anchor-free引用本文复制引用
基金项目
国家自然科学基金资助项目(62002253)
江苏省高校自然科学基金资助项目(19KJA230001)
出版年
2024
国家科技期刊平台
NSTL
维普
万方数据