Research On Object Detection Model Based On Convolutional Neural Network

The object detection model based on the convolutional neural network,with its excellent performance advantages,has shown great potential and value in the fields of transportation,security,medical care,etc.,and the significance of the detection model will be further deepened in the future.In recent years,the research on object detection models has also been developing rapidly,with new models and new structures appearing constantly,showing better detection results.This paper focuses on the shortage of the one-stage object detection model SSD in the insufficient testing accuracy and the design of small lightweight models,two models are designed respectively.The research work and innovation points of this paper are mainly in the following two aspects:(1)Improvements based on SSD.As a classic detection model,SSD has low accuracy nowadays,but its detection thought and structure are still classic.This paper proposes an improved version of SSD,MSSD(Modified SSD).In MSSD v1,a modified depthwise separable convolution module is proposed,which is used as a replacement for extra layers.In order to further improve the utilization of output features,adjacent features fusion is designed for reuse of features,the accuracy of MSSD v1 on the VOC2007 test set increased by 1.9%.The design of MSSD v2 is based on the backbone network and feature extraction structure.First,a type of VGG(Visual Geometry Group)convolutional neural network Rep VGG is used as the backbone network,and the original VGG-16 is replaced to obtain Better expression of basic features;After analyzing the advantages of dilated convolution and attention mechanism,a two-branch feature extraction module is designed that combines the advantages of the two together.At the same time,the new module uses a feature fusion module with the attention mechanism.The feature extraction module and the fusion module are embedded in Rep VGG,and a new model MSSD is obtained.The accuracy of the VOC2007 test set is 4.48% higher than that of SSD,the overall accuracy has a big improvement.(2)Design of a lightweight model.After analyzing the Shufflenet v2 structure,a lightweight and effective model Res-Shuffle is proposed based on its internal structure.The backbone network of Res-Shuffle is improved from the basic structure of Shuffle Net v2 which meets the needs of adapting to the object detection task.At the same time,the idea of identity mapping and dilated convolution is introduced,and the SSD detection process is followed.In addition,Res-Shuffle uses an adaptive positive and negative sample selection method.The introduction of this method has two advantages.On the one hand,the new model does not need to set numerous anchor boxes,which reduces unnecessary workload,on the other hand the selection basis of the sample is more reliable,and the quality of the sample is also guaranteed.In order to test the performance of Res-Shuffle and at the same time solve the task requirements of remote sensing aircraft object detection,this paper collects and organizes a special remote sensing aircraft image collection from multiple data sets.The experimental results show that Res-shuffle has achieved higher accuracy and speed at a lower parameter,and compared with other lightweight models it have a large progress.