Research On Multi-scale Object Detection Algorithms Based On Deep Neural Network

Object detection is a research hotspot in the field of computer vision,which realizes the location and classification of objects in images.Object detection are widely used in life,and objects in life are multi-scale.Different scale objects bring certain difficulties to detection.At present,most object detection algorithms use deep neural networks,but due to the deep network,the classification of small object is not satisfied.Therefore,in order to solve the problem of multi-scale object detection,improve the accuracy of small-scale object detection,and reduce the amount of calculation,this thesis mainly does the following work:(1)Firstly,the research background and significance of object detection are introduced,and the process of object detection based on deep neural network is briefly described.It is pointed out that multi-scale object detection is a common identification problem in life,and multi-scale object detection has certain difficulties.The last layer of deep neural network cannot detect the location of small object well.It ignores the small object.This paper proposes a feasible solution to some of the difficulties of multi-scale objects detection.(2)According to the feature map of deep neural network,there are different degrees of semantic information in different layers.The feature map of low convolution layer and the high convolution layer have different representations of object features at different scales.The paper proposes an algorithm called Concise Feature Pyramid Region Proposal Network which is for multi-scale object detection.It combines the image pyramid and the feature pyramid into a single model,and classifies and locates the objects of different scales in the appropriate convolution layer.After obtaining the classification prediction boxes,a certain fine-tuning location is performed.Finally a more accurate object location is obtained.Experiments show that the proposed method is effective for multi-scale object detection,especially for the detection of small objects.(3)In order to improve the detection accuracy of multi-scale objects further,this paper will combine multiple image pyramids and feature pyramids,and propose a CFPRPN algorithm based on model fusion.The small-scale object model and CFPRPN model will be merged to carry out multi-scale objects detection.The detection rate of small-scale targets is improved further,and effective results of multi-scale object detection are obtained,which improves the accuracy of multi-scale object detection.