基于逐级信息恢复网络的实时目标检测算法

Hierarchical Information Recovery Network for Real-Time Object Detection

随着卷积神经网络的发展,目标检测算法成为计算机视觉领域的研究热点,基于深度学习的实时目标检测算法需要同时兼顾检测精度和检测速度两项指标.不基于先验框的实时目标检测算法Center Net大幅提高了检测速度,但是由于其直接对低分辨率高层特征进行连续上采样,没有充分补充特征在下采样过程中丢失的空间细节信息,导致算法对目标定位不够准确,影响了检测精度.为解决这一问题,提出了一种基于逐级信息恢复网络(hierarchical information recovery network,HIRNet)的实时目标检测算法.该算法中,为对信息进行逐级恢复,设计了相邻层信息增强模块(adjacent layer information strength module,ALISM)和残差注意力特征融合(residual attentional feature fusion,RAFF)模块.通过构建ALISM模块,将中间层特征进行处理,分别为相邻层特征提供更多的空间细节信息和语义信息,提高低层特征的表达能力,输出更适宜进行信息恢复的特征.为进一步精确恢复损失的空间细节信息,HIRNet在上采样过程中逐级使用构建的RAFF模块,这一模块综合利用全局和局部注意力调整低层特征和高层特征的残差权重,再对两级特征进行加权融合,恢复高层特征在下采样过程中丢失的空间细节信息.在PASCAL VOC数据集和MS COCO数据集上的实验证明了所提算法的有效性.在MS COCO验证集上,HIRNet保证了检测的实时性,提升了算法检测性能,检测精度比Center Net算法提高了3.9%.

With the development of convolutional neural networks,object detection has become a focused research area in computer vision.Real-time object detection algorithms based on deep learning need to consider detection accuracy and speed.The real-time anchor-free object detection algorithm called CenterNet greatly improves the detection speed.However,it directly performs continuous upsampling of high-level features which are low-resolution.It does not fully recover the spatial details lost in the downsampling process,resulting in inaccurate positioning and low detection accuracy.To address this problem,a hierarchical information recovery network(HIRNet)is proposed.Here,the information is hierarchically recovered by developing an adjacent layer information strength module(ALISM)and residual attention feature fusion(RAFF)module.ALISM was designed to use the middle-layer features to provide more spatial details and semantic information for the adjacent layer features and improve the low-level features’discriminative power.Thus,its outputs were more suitable for information recovery.RAFF was hierarchically used in the upsampling process to further recover the lost spatial details.It used the global and local attention to adjust the residual weights of the low-level and high-level features,then fused the two-level features to recover the spatial details of the high-level features,which were lost in the downsampling.Experiments on PASCAL VOC and MS COCO datasets showed the effectiveness of the proposed algorithm.HIRNet guarantees real-time detection with an accuracy of 3.9% higher than that of the CenterNet on the MS COCO minival dataset,improving the detection performance.