Research On Ground Object Recognition In Aerial Images Based On Deep Convolutional Network

In recent years,UAV technology has been continuously developed and widely used in various industries,and object detection in aerial images is an important part of realizing these applications.However,due to various reasons such as the shooting angle and height of UAV aerial imagery,there are too many small objects in aerial imagery,and there is a phenomenon of scale variation.The existing aerial image object detectors face three main problems: 1)The dense connection method is used to achieve the fusion of low-level and high-level features to enhance the model’s ability to represent small object features,but this method is prone to feature redundancy;2)Introducing a non-local attention mechanism to learn more key information of objects,but the high consumption and low efficiency it brings will lead to difficulties in practical applications;3)The positive sample assignment process is biased towards medium and large objects,resulting in insufficient positive samples assigned to small objects.In response to the above challenges,this paper proposes a reasonable improvement strategy for the task of aerial image object detection.The main improvements are as follows:(1)In terms of feature fusion,this paper proposes a sparse multi-scale feature fusion method.In this method,two steps of interval-layer fusion and adjacent-layer fusion are used to realize the sparse fusion of features.Compared with the existing dense feature fusion methods,it can not only realize the effective fusion of shallow fine-grained information and deep semantics,but also avoid feature redundancy caused by dense feature fusion.After feature fusion,the network can effectively improve the feature representation ability of small objects.(2)As to attention mechanism,this paper proposes a double-shuffle attention mechanism.It can generate rich contextual information and refined feature output,which can enhance the attention to the target area while suppressing the attention to the background area.In addition,this mechanism can also improve the network’s ability to aware the object scale,which helps to expand the scope of model scale processing.It is worth noting that the attention mechanism subtly achieves a lightweight design through channel grouping and location block strategy to ensure the detection speed of the model.(3)In the aspect of positive and negative sample assignment,this paper proposes a novel anchor assignment strategy.In this strategy,the Io U value and confidence are comprehensively considered to provide a more reasonable metric for the quality evaluation of anchor.In addition,this method distinguishes small objects from medium and large objects during the assignment of positive and negative samples,which can effectively increase the number of positive samples assigned by small objects,thereby improving the contribution of small objects to the loss during network training.(4)In terms of model deployment,TensorRT and OpenVINO are used to accelerate the inference of the model to improve the real-time detection speed of the algorithm and facilitate the actual application and deployment of the algorithm.The experimental results show that the performance metrics AP of the aerial image object detection model designed in this paper is 29.7%,44.8% and 53.3% in the Vis Drone2019 dataset,the UAVDT dataset and the DIOR dataset,respectively.Compared with the benchmark YOLOv5 network,the performance metrics AP is improved by 5.6%(from 24.1% to 29.7%),5.5%(from 39.3% to 44.8%)and 2.9%(from 50.4% to 53.3%).The experimental results show that the method proposed in this paper can be effectively adapted to the task of aerial image target detection in a variety of scenarios.