Research On Lightweight Deep Model For Object Detection

With the development of deep learning and object detection models,the amount of parameters and computations of the models are increasing,and the requirements for hardware resources are getting higher and higher,which brings difficulties to the deployment of object detection models.Based on this,this paper conducts lightweight research on models in object detection scenarios,proposes effective model pruning and knowledge distillation methods,and verifies the advancement and effectiveness of the methods in experiments.The main work and innovations of this paper are as follows:(1)A model pruning method that fuses scaling factor and mutual information value ranking is proposed.Considering the BN layer scaling factor as the evaluation criterion of channel importance,the polarized L1 regular term is used to replace the L1 regular term for sparse training of scaling factors,and the important scaling factors and unimportant scaling factors are automatically distinguished,which is beneficial to the subsequent accuracy recovery training.Considering the mutual information values of channels and labels as the evaluation criteria of channel importance,the channel features are aggregated through global dual pooling,and the mutual information values of channels and labels are calculated on the experimental set.The final channel importance score is the ensemble result of scaling factor and ordering of mutual information values.The method in this paper prunes VGG-16,with an accuracy loss of 0.45%,which reduces the amount of parameters by 85.9% and the amount of computation by 58.2%without significantly reducing the network performance.The detection efficiency of the detector can also be significantly improved under various one-stage and two-stage object detection frameworks.The Cascade RCNN whose backbone network is Res Net101 only loses1.4 m AP,reducing 50.3% of parameters and 56.9% of amount of calculation.(2)We propose a knowledge distillation method based on instance screening and dualdomain attention.Considering that in the object detection scene,the positive and negative detection instances are not balanced and most methods rely too much on manually set parameters to balance,this paper proposes an instance selection module based on instance difference metric,which is automatically selected according to the detection results of the teacher model and the student model.Instances where distillation is most needed.In terms of feature distillation,feature distillation is performed on the filtered instances based on the spatial attention mask and the channel attention mask.Meanwhile,the spatial attention mask activated by the multi-head self-attention module and the channel attention mask activated by the SE excitation module are used to perform feature distillation on the global features of the backbone network.Considering the importance of attention map in feature distillation,the global attention map is distilled by L2 loss.In instance relation distillation,the Euclidean distance is used to measure the relation of instance features,and the distillation is performed by the L1 loss function.The experimental results show that the knowledge distillation algorithm in this paper can significantly improve the detection effect of the student model,and the effectiveness of each module has also been proved in the ablation experiment,and the method is insensitive to hyperparameters and has strong robustness.(3)The knowledge distillation method is applied to the accuracy recovery training after model pruning,and the auxiliary fine-tuning method is used to perform the accuracy recovery training.A dynamically decaying distillation loss term coefficient is proposed to ensure that the distillation strategy can bring better accuracy recovery to the pruned model without affecting the upper limit of the model.The effectiveness of the method is verified under various mainstream object detection frameworks.On Faster RCNN whose backbone network is Res Net101,m AP is increased by 0.6,reducing 49.5% of parameters and 56.3% of computation.On Retina Net,whose backbone network is Res Net101,m AP is increased by 0.3,reducing 51.7%of parameters and 59.4 % of the calculation amount.Experiments are carried out in the air-toground UAV detection scene with high requirements for detection speed.The results show that the model pruning algorithm integrating the knowledge distillation strategy can reduce the number of parameters of the model to a greater extent,improve the detection speed,and improve the detection speed.Both small targets and multi-targets can achieve good detection results.