Research On The Small Object Detection Algorithm Based On Feature Fusion

There are a large number of small objects in real-world object detection scenarios,such as unmanned driving and intelligent surveillance.Due to the small size,incomplete contextual information,random number and distribution of small objects,the detection ability of generic object detection algorithms is not ideal.A feature fusion method is used by researchers to extract and optimally combine different features from the same pattern,which can improve the ability of models to learn multi-scale features.In this thesis,small object detection is investigated by combining the feature fusion method as follows:(1)The feature expression capability of small objects is weak,and the generic object detection algorithm does not sufficiently extract shallow features,resulting in unsatisfactory detection accuracy of the model.To solve this problem,this thesis proposes a small object detection network based on skip connections with the shallow feature.First,a upsampling method called deconvolution is used by the model to adjust the size of the feature map.Then a dense skip connection structure is introduced using feature fusion method of feature addition to fuse shallow features of different scales.The skip connection can reduce the loss of small object features in the forward transfer process and maximize the retention of small object-related detail information,thus enhancing the representation of small object features.Finally,some experiments are conducted on the public datasets,and the results show that the improved model can improve the detection ability of small objects.(2)The deep feature map contains mainly high-level semantic information,from which the model can hardly obtain the features related to small objects.To address the problem,a small object detection network is proposed based on adaptive spatial feature fusion in this thesis.First,the model builds a feature pyramid module in the middle layer of the network based on skip connections with shallow feature to fully fuse feature information in feature maps at different layers,which improves the ability of the model to learn multi-scale features.Then the adaptive spatial feature fusion module is constructed on the feature pyramid module,which learns the weights of each feature adaptively according to the reliance of different scale objects on different information.In addition,the shortcoming of the confidence calculation method for small object detection is investigated and improved in this thesis.Finally,some experiments are conducted on the public datasets.Compared with other models,the experimental results show that the model has better performance for small objects.