Research On Lightweight Small Target Detection Algorithm Based On UAV Image

In recent years,unmanned aerial vehicle(UAV)technology and artificial intelligence algorithm are increasingly mature,and the application scope of UAV in the field of transportation is constantly expanded.The combination of UAV and object detection algorithm can further promote the development of intelligent transportation system.In view of the characteristics of small object and limited computing capacity of UAV airborne equipment,this paper studies the small object detection algorithm and lightweight network design.The main research contents are as follows:(1)In view of the small size and dense distribution of ground objects taken by UAV at high altitude,this paper proposes a small object detection network RSOD based on multiscale feature learning.Firstly,the shallow features with more fine-grained information are used to assist the localization and recognition of small objects.Then,the local and global information around the object is extracted through the spatial pyramid pooling layer,and it is combined with the shallow and deep information in the feature pyramid to further enrich the feature expression of small objects.Then,the importance of each output feature of the feature pyramid is measured by adaptive learning,and weighted fusion is carried out to enhance useful features and suppress useless features.Finally,the sensitivity of the model to feature channels is improved by the improved SE attention mechanism,and the performance of small target detection is further improved.(2)In terms of lightweight network design,this paper introduces a scale factor for each channel in the RSOD network to measure the importance of the channel.Then regularization is added in the training process,and the scale factor is sparsely trained.Next,the channels with the scale factor close to zero are pruned and all the input and output connections and corresponding weights are deleted.Finally,the pruned model is retrained and the model accuracy is restored by fine-tuning.After several pruning and fine-tuning,the model size and small object detection accuracy reach a good balance.(3)In order to simplify the acquisition process of lightweight network and pursue fewer model parameters,faster running speed and higher small-target detection accuracy,this paper uses the depth-wise separable convolution and attention mechanism to construct the lightweight feature extraction network,and instead of the traditional idea of anchor-based framework,each point in the feature map is mapped back to the original image,and the bounding box is directly regressed,which simplifies the detection process.Finally,the dilated convolution is used to obtain the features of different receptive fields,and the feature pyramid network is injected to supplement the context information.Meanwhile,channel and spatial feature refinement mechanisms are introduced to suppress the formation of conflict information in multi-scale feature fusion and improve the detection accuracy of small objects.