Research On Small Object Detection Based On UAV Images

With the development of data collection technology on Unmanned aerial vehicles(UAV),object detection algorithms based on UAV images have gradually become one of the research hotspots in computer vision.Compared with large objects in natural scenes,objects in UAV images are usually small due to their distant perspective,and there are challenges such as weak features and dense distribution.To address the difficulties and challenges in small object detection in UAV images,this paper proposes a small object detection algorithm based on multi-scale feature fusion and a small object detection algorithm based on anchor-free regression mechanism.The main innovations in this paper include the following two aspects:(1)To address the challenge of weak features in small object detection from UAV images,we propose a small object detection algorithm based on multi-scale feature fusion.Firstly,to enhance the attention of the feature extraction network on small objects,we introduce an Enhanced Spatial Pooling Module(ESPM)that enhances the feature extraction capability by embedding the attention mechanism into each pooling layer of the spatial pyramid.Secondly,to further fuse shallow detailed features with deep abstract semantic features,we construct a Multi-Scale Feature Pyramid Network(Ms FPN),which promotes the network to extract more discriminative feature representations by introducing shallow information and weighting the features of different layers.In addition,to address the issue of misaligned classification and regression features,the algorithm introduces De Coupled Head(DHead)to separate two tasks and learn the features that are more relevant to each task.Finally,to improve the slow anchor box regression speed in the training process of anchor box algorithms,a loss function with vector angles(SIo U)is adopted to calculate the angle vector between the predicted box and the true box,and improve the model inference speed.Experimental results on the Vis Drone2019 UAV image dataset demonstrate the effectiveness of the proposed algorithm,which outperforms the baseline algorithm YOLOv5 with an average precision improvement of10.0% in small target detection.(2)To address the issue of complex anchor settings and high computational cost in anchor-based object detection models caused by dense target distribution in drone images,this paper proposes a small object detection algorithm.Firstly,the algorithm embeds Contextual Transformer Block(Co T Block)based on the self-attention mechanism into the feature extraction network,which assists in improving the representation ability of the model by mining contextual information.Secondly,we designed a simple Bidirectional Fusion Pyramid Network(BFPN)to enhance the semantic interaction between shallow and deep features in the channel dimension.Furthermore,a label expansion strategy and a quality-based loss function are proposed to address the problem of imbalanced positive and negative samples,which expands the positive sample range of the center point of the bounding box and uses the bounding box quality score as a label to guide the training and testing of the classification branch,promoting the model to fully mine high-quality samples.Experimental results show that the proposed model achieves a 6.2% average precision improvement compared to the baseline algorithm FCOS on the Vis Drone2019 dataset.