| With the rapid development of computer technology,many target detection algorithms based on deep learning have been proposed,which have become a popular research direction in computer vision and are applied in practical scenarios.Among many targets,small targets are small in scale and occupy less pixel information.These characteristics make the feature extraction process of small targets more difficult,resulting in lower detection accuracy.Therefore,how to improve the accuracy of small target detection has become a research difficulty in the field of target detection.This thesis selects SSD(Single Shot Multi Box Detector)as the basic research algorithm,and the main research work is as follows:Aiming at the problem that the algorithm does not establish the semantic relationship between each output feature layer,a small object detection algorithm that introduces a topdown feature fusion path is studied.In order to enhance the feature extraction ability of the model for small objects,this thesis establishes a feature fusion path from the deepest feature layer to the shallow feature layer on the basis of repeatedly extracting the feature information of the shallow feature layer by using the multi-scale feature extraction network.Enhancing the semantic information of shallow feature layers for small object detection.Comparative experimental results show that compared with the original SSD algorithm,the object detection accuracy and small object detection accuracy of this method on the PASCAL VOC dataset are increased by 3.3% and 5.7%,respectively.In order to further improve the model’s detection performance of small objects in natural scenes,this thesis adopts a small object detection algorithm that integrates vision mechanism and multi-scale features and attention mechanism.First,the visual mechanism module is used to increase the receptive field of the shallow feature layer.Then,a multi-scale feature fusion network and a deep feature enhancement module are used to enhance the semantic information of the shallow feature layer and the deep feature layer,respectively.Finally,the attention mechanism module is used to enhance the learning ability of key information of small targets.Compared with the original SSD algorithm,the object detection accuracy and small object detection accuracy of this method on the PASCAL VOC dataset are increased by 3.9% and6.6%,respectively,which effectively improves the detection performance of the model for small objects.In addition,in order to further optimize the ability of the shallow feature layer to extract the feature information of small objects,this thesis also adopts a small object detection algorithm that fuses the backbone enhanced semantic information for detection.First,in order to obtain multi-scale feature information,the feature information of different feature layers of the shallow backbone network is fused through the backbone feature fusion module.Secondly,the backbone feature enhancement module is used to enhance the semantic information of the backbone network layer,and the feature fusion module is used to splicing it with the backbone feature layer.Finally,the output feature maps of each feature layer are passed through an efficient channel attention module,which aims to strengthen the training of small target detail feature information.The experimental results show that the object detection accuracy and small object detection accuracy of this method on the PASCAL VOC dataset are 4.5% and6.6% higher than the original SSD algorithm,respectively,which proves that the model can significantly improve the detection ability of small objects.The research results of this thesis will help to improve the performance of small target detection in natural scenes,so that this technology can be better used in production and life. |
