Research On Acceleration Method Of Object Detection With Convolution Neural Network In Mobile Sense

With the development of computer technology and the general trend of the development of Internet of Things and artificial intelligence,object detection technology,as an important branch of computer vision,has been widely applied to face recognition,video surveillance,vehicle assisted driving,and mobile robots and other fields.At present,there are many methods for processing object detection,which are roughly divided into two categories:image processing based detection methods and deep learning detection methods based on convolutional neural networks.However,because the image processing-based detection method has low accuracy and cannot meet people's increasing detection accuracy requirements,it has gradually been replaced by the latter.Since the computational process of the convolutional neural network requires a large amount of computational resources and a large amount of model storage,it poses a huge challenge to the deployment on the mobile side,such as mobile phones and mobile robots,which are inconvenient to deploy large-scale computing servers.Therefore,it is of great significance to study the detection task acceleration and compression strategy based on convolutional neural networks.This thesis mainly studies the object detection technology of mobile end,and proposes a method of model compression,model cropping and model quantification,aiming at solving the problem that mobile computing is not strong and storage capacity is limited.This paper mainly studies in the following aspects:1.Research on compression method based on network structure model.Research and compare several existing deep learning-based feature extraction network architectures,and propose an improved model architecture based on separable convolution.The model architecture can be reduced by the form of group convolution to ensure the accuracy of the model.The parameter quantity of the model and the calculation amount of forward propagation are small,which improves the speed of object detection.2.Research on tailoring algorithms for specific scene models.Aiming at the general model architecture,there are a large number of parameter redundancy problems in a specific scenario.By counting the location information of the redundant parameters,the clipping is given under a certain threshold to achieve further compression of the model parameters and acceleration of the forward propagation.3.Research based on one-bit network quantization algorithm.Using a computer to calculate a 32-bit floating-point operation in one clock cycle can calculate the characteristics of the 64-bit operation,quantify the floating-point operation of the traditional model into a bit type,and further accelerate the forward propagation of the model.