Small Object Detection Based On Intra-Inter Layer Fusion

With the development of deep learning technology,the object detection based on deep learning has made significant progress.However,the small object has small imaging area against complex background,it is difficult to extract high-quality features,resulting in difficulty in small object detection.At present,the deep learning based object detection framework is divided into three categories: single detection layer structure,multi-scale detection layer structure and multi-layer feature fusion structure.For small object detection,the object detection framework with the single detection layer structure uses deep-level features to detect small objects,but the feature information of small object is seriously lost and the feature receptive field does not match the scale of the small object.The object detection framework with the multi-scale detection layer structure uses shallow-level features to detect small objects,but its expression ability is insufficient.The object detection framework with the multi-layer feature fusion structure fuses the multi-layer features together.Then deep-level semantic information can flow to shallow-level layer,and shallow-level layer is used to detect small objects.However the correlation between multiple feature channels are not utilized,the expression ability is still insufficient.In this thesis,on the basis of the object detection framework with multi-layer feature fusion,the small object detection framework based on intra-inter layer fusion is proposed.The main research is as follows:(1)In the multi-layer feature fusion based object detection framework,the existing detection algorithm initializes the preset box according to experience.When detecting small objects,the preset box initialized by experience is quite different from the real object,so it is not stable enough at the beginning of training.In this thesis,three improved algorithms of classical K-means are analyzed.According to the characteristics of clustering tasks,the K-means++ algorithm is adaptively modified,then the dimension of width and height of preset box is clustered.The initialization method is closer to the real object,which makes linear regression model training and the preset box fine-tuning more reasonable and stable,and improves the detection effect.(2)In the multi-layer feature fusion based object detection framework,the batch normalization is commonly used to normalize network.The multi-layer feature fusion network has high complexity.In the actual training process,the batch size is set smaller.The mean and variance calculated by the BN algorithm lack statistical characteristics,so the normalization result is poor.In this thesis,the GN algorithm is used to normalize the network,which does not rely on the batch size,it can get good normalization result when the batch size is small.(3)In the multi-layer feature fusion based object detection framework,the multi-layer feature fusion network structure explores the inter-layer relationship between multiple feature layers.In this thesis,the correlation between multiple feature channel is further analyzed from the qualitative and quantitative perspectives.The channel fusion module is designed to construct the intra-inter layer fusion network.The network not only enhances the expressive ability of the output features by using the intra-layer relationship between the feature channels,but also gradually integrates the deep-level layer with strong expression ability into the shallow-level layer by using the inter-layer relationship.The network uses shallow-level features to detect small objects and can extracts high-quality small object detection features.(4)The simulation results show that the proposed object detection framework based on intra-inter layer fusion has achieved a mAP of 83.7% on the aerial vehicle dataset,which is much better than that of Faster RCNN and RFCN and slightly better than that of SSD.The proposed object detection framework also has a significant improvement over the RON and RefineDet,and its detection speed is within an acceptable range.