基于模型剪枝和知识蒸馏的船舶目标检测方法

Ship Target Detection Method Based on Model Puning and Knowledge Distillation

针对基于深度学习的目标检测技术需要大量计算、卷积网络里大量参数占用内存空间导致模型太大无法较好进行部署等问题,提出一种基于模型剪枝和知识蒸馏的船舶目标检测方法.首先构建小样本船舶数据集,通过数据增广来扩充数据集数量,然后采用YOLO v3模型对其进行训练,接着对网络进行模型剪枝,最后采用知识蒸馏技术对剪枝后的模型进行蒸馏,补偿剪枝后模型的检测精度损失.试验表明,剪枝后压缩了模型体积,且仍有较高的检测精度.

At present,target detection technology based on deep learning needs a lot of calculation,and convolution network has large numbers of parameters that occupy memory space,so the model is too large to be deployed well.This paper proposes a ship target detection method based on model pruning and knowledge distillation.Then,YOLO v3 model is used to train it,and the network model is pruned.Finally,knowledge distillation technology is used to distill the pruned model to compensate for the loss of detection accuracy.Experiments show that the pruned model not only has the advantage of model size,but also has high accuracy in detection accuracy.