知识蒸馏和特征融合相结合的目标检测算法

Research on Target Detection Algorithm Based on Knowledge Distillation

目前Faster Rcnn是主流的目标检测框架之一,检测精度提升的同时伴随着更深的特征提取网络,但带来的大量的参数和计算开销使得这些算法难以应用在对存储空间和参数量有要求的移动设备上。为了在降低模型复杂度的同时保持与复杂网络相近的性能,论文将知识蒸馏方法用在目标检测框架中的特征提取网络,且为更好提升浅层特征提取网络性能,在知识蒸馏阶段引入了特征融合技术。网络规模相同的情况下,使用该方法的特征提取网络的检测精度比没有经过知识蒸馏的特征提取网络的检测精度高了6.53%。保证检测速度提升的同时,经过蒸馏后浅层网络的精度与复杂网络的精度相差不多,这证明了该方法的有效性。

At present,Faster Rcnn is one of the mainstream target detection frameworks.The improvement of detection accuracy is accompanied by a deeper feature extraction network,but large number of parameters and additional computational overhead make it difficult to apply these algorithms to storage space and parameter requirements on mobile devices.In order to reduce the com-plexity of the model while maintaining performance similar to the complex network,this paper uses the knowledge distillation method in the feature extraction network of the target detection framework.In order to better improve the performance of the shallow fea-ture extraction network,feature fusion technology is introduced in the knowledge distillation stage.In the case of the same network scale,the detection accuracy of the feature extraction network using this method is 6.53%higher than that of the feature extraction network without knowledge distillation.While ensuring the increase in detection speed,the accuracy of the shallow network after distillation is similar to that of the complex network,which proves the effectiveness of the method in this paper.