Research On Deep Learning Networks For Small Object Detection Based On Multi-level Feature Fusion

In recent years,convolutional neural networks have achieved better performance in the field of computer vision,while their expanding application scenarios have also generated some more challenging special tasks like small object detection.Since small target detection puts forward higher requirements on the feature extraction and fusion,mainstream algorithms for general detection tasks cannot meet the needs of small target tasks.Consequently,this thesis proposes our improvement method to create a dedicated network for small target detection.Visualization analysis shows that the shallow layer of the common detection network under the small target task has the problem of low feature quality and partial feature redundancy,while the deep layers of the network have the problem of ignoring important details and lacking necessary semantic information.In view of the lower feature quality and connection redundancy of the shallow layers,this thesis first adds a convolution branch with a smaller receptive field to extract more detailed texture information.Secondly,we take the qualitative analysis results of the common detection networks' feature fusion strategies into account and propose the Efficient Dense Block.We further apply the idea of shallow feature fusion to the module-level connection,and build a new backbone network where shallow features are reused across modules.Aiming at the problems of ignorance of detailed information and lack of semantic information,this thesis introduces and enhances the channel-spatial attention module.We design an innovative multi-frequency domain fusion method to replace the traditional Global Average Pooling,with the purpose of emphasizing important shallow features,thereby increasing effective attention to small target areas.The spatial attention module in this thesis uses different scales of dilated convolution to improve the network's ability to distinguish background areas and target areas.At last,this thesis chooses small target detection in the images from Unmanned Aerial Vehicles(UAV)as the task,and conducts comparative experiments with the general detection network and some small object detection algorithms.The experiments show that our improved backbone network has achieved an accuracy increase of about 8% in small target tasks compared to the general network.Moreover,we present both the detection results and heat map of our proposed attention module to visually verify the mitigation of the problem about multi-level feature fusion.