基于优化RetinaNet的自适应特征融合的轻量化目标检测方法

Lightweight Target Detection Method Based on Adaptive Feature Fusion of Optimized RetinaNet

针对目标检测算法中计算量大、模型复杂等问题,造成在嵌入式平台计算资源有限的应用场景下难以部署的现象,提出一种基于优化RetinaNet的自适应特征融合的轻量化目标检测方法。提出的算法参考了GhostNet中的Ghost Module模块以减少模型参数量。通过一种空间特征融合机制,提高特征的尺度不变性。融合了结构重参化的思想,增加训练深度,实现多分支训练,单分支推理,更好地提升模型的推理性能。提出的方法在PASCAL VOC2007和COCO两种常用的目标检测数据集上进行评估,平均精度为54.1%,优于RetinaNet的平均精度,实验结果表明,提出的方法推理时所占的内存为170.71MByte,是RetinaNet所占内存的44.27%,表明提出的算法在保证精度的前提下极大提高网络的推理速度。

Aiming at the problems of large amount of computation and complex model in target detection algorithm,which make it difficult to deploy in application scenarios with limited computing resources on embedded platform.The lightweight target detection method based on optimized RetinaNet is proposed with adaptive feature fusion.Firstly,the proposed algorithm refers to the Ghost Module in GhostNet to reduce the number of model parameters.By means of a spatial feature fusion mechanism,the scale invariance of features is improved.Secondly,the idea of structural reparameterization is integrated to increase the depth of training,realize multi-branch training,single-branch training,and better improve the detecting performance of the model.The method is evaluated on two common target detection datasets,PASCAL VOC2007 and COCO.With an average accuracy of 54.1%,better than that of RetinaNet.The experimental results show that the memory taken by the proposed method is 170.71MByte,which is 44.27%of the memory taken by the RetinaNet,indicating that the proposed algorithm can greatly improve the network inference speed without ensuring the accuracy.