Research On Compression Technology Based On Object Detection Network

In the past few years,computer object detection technology has developed rapidly.Object detection algorithm refers to a given target image,through which specific objects in the image can be detected,that is,to achieve both localization and classification purposes.It is currently one of the research hotspots in the field of computer vision.With the development of 5G network commerce and urban modernization,the country attaches great importance to the construction of key directions such as intelligent transportation and public safety.As a necessary basic research technology,object detection algorithms play a key role in real-world scenarios such as autonomous driving,urban management,and disaster relief.Although most current object detection models have achieved excellent performance in detection,in resource-constrained environments such as mobile devices,embedded platforms,and the Internet of Things,most deep neural networks require significant computational resources.At the same time,its slow training and reasoning speed and huge space occupation make us have to focus on lightweight object detection networks.Lightweight object detection networks can reduce computational complexity and space consumption,and have broad application prospects in various resource constrained or demanding application scenarios such as computing speed,energy consumption,and storage.At the same time,there are also new challenges in how to propose effective model training schemes so that lightweight object detection models can still maintain excellent performance.In order to solve the above problems,our main contribution is to propose a new feature extraction network called CompleNet.We designed the network as a parallel structure to accelerate the feature extraction process and optimize feature fusion between different branches of the module.In subsequent improvements,we added the Squeeze and Extraction Networks attention mechanism module before the final output of the feature map from this module to improve the feature extraction effect,while introducing pattern parameters λ,The detection performance of the network can be quickly adjusted for different detection targets.In addition,compared to current mainstream feature extraction modules,this module has significant advantages in parameter size and training speed.Compared to other feature extraction networks such as ResNet50,it is particularly suitable for deployment on mobile terminals or embedded devices.At the same time,the CompleNet feature extraction network has also achieved better performance in reasoning time,increasing by 1.477 times compared to the ResNet50 feature extraction network.In addition,in order to solve the problem of performance degradation after model lightweight,we proposed a knowledge distillation training method specific to the CenterNet object detection network-multi teacher joint knowledge distillation scheme.This scheme can effectively solve the problem of performance degradation caused by model lightweight,greatly reducing the performance gap between teacher models and student models.Using large-scale complex models as teacher models to teach small,lightweight models as student models for training.Compared to directly trained lightweight object detection networks,using this knowledge distillation scheme can achieve better detection performance after the same number of training rounds.In subsequent experiments,we innovatively proposed a distillation attention mechanism to optimize the training effect of multi teacher joint knowledge distillation,using the difference in the loss between different teacher networks and real tags.On the VOC2007 dataset,taking the MobileNetV2 lightweight network as an example,compared to the traditional CenterNet(the backbone network is ResNet50),the parameter size has been reduced by 74.7%,and the inference speed has been improved by 70.5%.On the mAP,only 1.99%has been reduced,achieving a better "performance speed" balance.After 100 rounds of training,the multi teacher combined knowledge distillation training method has been used,which improves the mAP by 11.30 compared to the ordinary training method.