Research And Application Of Improved Algorithm For Small Object Detection Based On YOLO

The object detection task is to detect the position of the object in the visual image and judge its category.In recent years,deep learning has achieved great success in the field of machine vision,and object detection technology has also achieved excellent results,and has been well applied in many fields,such as industrial manufacturing,autonomous driving,intelligent robots,and medical diagnosis.However,since small objects occupy a small area in the image and are easily disturbed by the background,the detection performance of deep learning object detection algorithms for small objects is still not ideal.Therefore,this thesis starts with the research of small object detection,analyzes the evaluation index and adaptability of small object detection,and improves the existing object detection algorithm to improve the performance of small object detection.The main research content of the thesis is as follows:(1)Aiming at the serious problem of feature information loss after multiple downsampling of the features of small objects in the deep convolutional neural network,we studied the introduction of more shallow feature maps in the feature fusion part of YOLOv4 for mutual fusion;at the same time,in order to make full use of multi-scale features,a mixed attention mechanism is introduced to strengthen important features to improve the network’s attention to the feature information of small objects,so that the model can locate and identify small objects more accurately.Finally,the effectiveness of the improved method was verified experimentally on the Vis Drone2019 dataset and the PASCAL VOC dataset.The accuracy of the improved algorithm increased by 3.38%on the Vis Drone2019 test set and 0.81% on the PASCAL VOC test set.(2)Aiming at the problem of small object detection data sets in specific scenarios,a method based on Poisson fusion is proposed to fuse target samples and background images to achieve the purpose of expanding image data sets;at the same time,the target samples are fused with the background and recorded the location of the object sample in the image to generate the corresponding label,thereby saving the cost of manual labeling and avoiding errors caused by manual labeling.The experimental results show that the accuracy of the algorithm trained with the expanded data set is 43.69% higher than that before the expansion.(3)In order to deploy the small object detection algorithm to the mobile terminal,the Efficient-YOLOv4 lightweight small object detection model is proposed to solve the problem of limited computing power and memory space on the mobile terminal,using EfficientNet images with low parameters,low calculation and high performance perform feature extraction and build a feature pyramid with EfficientNet.Through comparative experiments,it is found that the accuracy rate of the improved algorithm is2.55% higher than that before the improvement,the model size is reduced by 191.5MB,and the detection speed is increased by 10 frames per second.