The Object Detection Method Based On Lightweight Deep Neural Network

Object detection is an important application area of computer vision and image and video processing technology.Its task is to locate and classify object entities in images or videos based on their visual features.With the application of deep learning technology in the field of computer vision,the performance of object detection models has been greatly improved.However,in practice,most models are difficult to deploy at the edge due to the large amount of computation and parameters,so they are lightweight.The research and application of object detection algorithms are becoming more and more important.Designing a lightweight object detection network requires consideration of the application scenarios,because such models are usually deployed on mobile devices.First of all,it is necessary to flexibly adjust the model size according to the difference in computing power of the devices? secondly,when considering the platform compatibility of the lightweight object detection model,it is required to compress the model without changing the network structure? thirdly,when considering the storage space and power consumption overhead However,there are a large number of floating-point operations in general deep learning models,which are not suitable for running on mobile devices directly.Therefore,in response to the above-mentioned problems,the work carried out in this thesis is as follows:(1)In view of the fact that the lightweight object detection model cannot flexibly adapt to devices with different computing power,from the perspective of optimizing the network structure,the channel pruning method is used to compress the object detection model,and by reducing the correlation of the BN layer during training iterations,the pruning is improved.branch accuracy.For mobile devices with different performance,the ratio of network pruning can be adjusted according to their computing power,which improves the flexibility of model deployment and enables the model to adapt to a variety of different mobile devices.(2)The lightweight object detection model has special requirements for the system software and hardware environment.From the perspective of not changing the basic structure of the model,the knowledge distillation method is used to improve the object location and category accuracy output by the lightweight object detection model,and through the domain The aligned method performs feature spectral cuedistillation to improve the features of the student network.The distilled lightweight model can be directly deployed in the original system without the need for hardware and software to optimize it.(3)Aiming at the problem of large storage and computational overhead of floating point numbers in lightweight object detection network,starting from the storage form of object detection model parameters,a method of quantizing network parameters through fixed points is proposed,using perceptual quantization training and then parameter quantization,and by introducing soft The quantization function allows the quantization error to participate in training.Converting the original object detection model from floating-point computation to fixed-point computation can reduce the power consumption of the processor and prolong the working time of the mobile device.