Small Object Detection Based On Deep Convolutional Neural Networks

Object detection is widely used in security monitoring,biomedical and industrial defect detection and other fields,of which small object detection is the focus and difficulty of detection.Although great progress has been made in detecting small targets using deep convolutional neural networks,there are still problems such as low detection accuracy and high missed detection rates.Therefore,based on the deep convolutional neural network,this paper conducts research and analysis on the small object detection algorithm,and specifically studies the following aspects:(1)In view of the current situation of domestic and foreign research on object detection,the convolutional neural network and several deep learning target detection algorithms were studied,the advantages and disadvantages of various algorithms and the scope of application were analyzed and compared,and finally starting from the detection speed and accuracy,YOLOv3 was established as the research basis,and the small object detection model was developed to adapt to the needs of small object detection.(2)Aiming at the problems of low detection accuracy and high leakage rate of YOLOv3 small targets,an improved and enhanced model of YOLOv3 is proposed.First,the hybrid attention mechanism CBAM is introduced to enhance the network’s extraction of small target features.Secondly,based on RFB＿Net designed a cascaded RF feature enhancement module,and proposed two different small object detection models by embedding different positions in the model structure.Finally,combined with the preheating and cosine annealing algorithm of the learning rate,YOLOv3 is constructed using Pytorch and trained for verification.Select pictures on the network to test the improved model for object detection,and conduct research and analysis of the test results.(3)Aiming at the problem of weak computing power of edge devices and mobile devices,an improved model of lightweight YOLOv4-tiny is proposed.Design and increase SPD module,improve network width and feature extraction capabilities,redesign CSP module,reduce computational complexity,introduce channel attention mechanism SE,and add feature enhancement module after lightweight design.Ablation experiments are carried out on all improved models,and the experimental results are evaluated and analyzed to verify the effectiveness of the improved algorithm model.(4)According to the actual deployment of the algorithm model,the algorithm is deployed to the mobile platform for actual target detection testing with ROS mobile robot as the carrier,the detection effect of the actual deployment is evaluated,and the final experimental detection effect is analyzed and studied.Figure 73 Table 6 Reference 83...