针对不同尺度的目标容易造成模型的特征提取和尺度变化的适应性问题,提出一种融合自适应感受野的目标检测算法.通过对特征图采用拆分、卷积和融合的方式构造通道自适应感受野模块,以提取不同尺度的感受野,提高对尺度变化的适应能力;在网络结构中引入通道自适应感受野模块和RepVGG模块,采用FC模块(concatenate with convolutions)过滤冗余特征,强化模型的特征提取能力;采用Alpha-CIOU损失和知识蒸馏优化训练,提高算法的检测能力.在Pascal VOC和MSCOC O数据集上的实验结果表明,该算法在尺度变化、精度和速度等方面取得了优秀的性能.
Abstract
To solve the problem that objects of different scales are easy to cause the feature extraction of the model and the adap-tability of scale change,an object detection algorithm based on adaptive receptive field was proposed.The channel adaptive recep-tive field module was constructed using splitting,convolution and fusion on the feature map to extract receptive fields of different scales and improve the adaptability to scale changes.The channel adaptive receptive field module and RepVGG module were in-troduced into the network structure,and the FC module(concatenate with convolutions)was used to filter redundant features to enhance the feature extraction ability of the model.The Alpha-CIOU loss and knowledge distillation were used to optimize the training to improve the detection ability of the algorithm.Experimental results on Pascal VOC and MS COCO datasets show that the proposed algorithm achieves excellent performance in terms of the scale change,accuracy and speed.
维普
万方数据