Research On Small Object Detection Method Based On Machine Vision

With the development of science and technology,the computing power of computers has been greatly improved,which has promoted the rapid development of deep learning.Therefore,breakthrough progress has been made in the field of machine vision object detection in recent years.With the increasing maturity of object detection technology,object detection has also begun to be applied to various fields such as medical treatment,security,and autonomous driving.Although the detection accuracy of object detection for large and medium-sized objects has been very high,the detection accuracy for small objects is still very low.The reason is that the features of small objects are not obvious,and it is difficult for the existing neural network to extract effective features for detection.Therefore,it is a challenging research to improve the feature extraction ability of neural networks for small objects and improve the accuracy of the algorithm for small object detection.This paper is based on the application scenarios of driverless cars,and studies the shortcomings of small object detection at the current stage.The contributions and innovations of this paper are as follows:In order to arrive the real-time detection requirements of unmanned driving scenes,this paper studies the existing one-stage object detection methods and proposes an improved model Bi-YOLO based on YOLOv4.Improve the structure of the Neck part in YOLOv4,use Bi FPN to fuse the features in the feature extraction network,so that the generated feature map has rich detailed features and semantic features,thereby improving the accuracy of small target detection;and using Bi-YOLO model and the existing one-stage object detection algorithm model is compared with experiments to verify the accuracy of the improved algorithm in small object detection.In view of the limited computing power and memory space of mobile devices,in order to improve the feasibility of deploying small object detection algorithms on mobile devices.This paper proposes a Ghost-YOLOv4-Tiny lightweight small object detection model based on YOLOv4-Tiny.The improvement is to use the Ghost module to upgrade the CSPDark Net backbone network and replace the CSP module structure.At the same time,in order to reduce the amount of calculation and reduce the loss of detection accuracy,the features are fused with reference to the structure of the feature pyramid.Through comparative experiments,it is found that this method reduces the calculation amount of the algorithm and the file size of the model under the condition that the speed and accuracy of model detection remain unchanged.In order to verify the effectiveness of the improved lightweight network in this paper,the lightweight network model is deployed on a low-speed campus cruise unmanned vehicle,and the detection effect of the lightweight network in practice is verified through actual testing.