Research And Application Of Object Detection Model Compression

In recent years,with the rapid development of deep learning,Convolutional Neural Networks(CNNs)have become the mainstream method for various computer vision tasks such as image classification,object detection,semantic segmentation,and depth estimation.Current deep learning-based object detection models often utilize large convolutional neural networks to extract features for higher accuracy.These models require high-end GPU chips to achieve real-time computing.This makes it difficult to deploy object detection models on edge devices with limited computational resources.To solve the problem of running deep learning models on devices with limited computational resources,a large number of model compression methods have been proposed,of which model pruning is one of the main directions.Traditional research on model pruning is mostly focused on image classification tasks.For object detection model pruning,a common practice is to use pruned CNNs as the backbone network(feature extraction network)of the object detection model to achieve the purpose of lightweighting.However,the backbone networks obtained using this approach are not optimized for the target detection task,and it is difficult to achieve the desired results.In this paper,the channel pruning algorithm is directly applied to the object detection model,and a layerwise adaptive channel pruning strategy is proposed in combination with sparsification training based on polarization regularizer.The method considers the differences in the sparsification results of different layers in the network,and uses adaptive pruning thresholds for different layers separately,thus avoiding unreasonable results in pruning.Meanwhile,the layer pruning method based on the layer importance estimation algorithm is proposed on the basis of channel pruning for the limitation of channel pruning in reducing the inference latency of the model.The method uses the channel pruning results of different layers to calculate their sparsity,and uses the sparsity to estimate the importance of the layer,thus providing a basis for layer pruning.In this paper,channel pruning and layer pruning are combined to form a joint pruning algorithm.In this paper,experiments based on the YOLOv5 series object detection model are conducted on the UAV target detection dataset Vis Drone-DET2019 and the remote sensing image dataset DIOR.The results of the experiments show that the pruning algorithm proposed in this paper can obtain better models than the handdesigned ones with similar number of parameters,which verifies the effectiveness of the method.In addition to studying the pruning of object detection models,a compression framework for object detection models is proposed in this paper.Common sparsification algorithms,model pruning algorithms,generic model thinning methods and model export are integrated in this framework.The whole compression process is driven and managed by a unified configuration file,which is easy for engineers to debug and reuse in more models.At the same time,the framework provides an extensible interface for users to customize new algorithms.The compression framework proposed in this paper bridges the entire process from object detection model compression research to edge computing device deployment.