Research On Small Target Detection Algorithm For Aerial Photography Images Based On Deep Learning

With the booming development of artificial intelligence,computer vision-based target detection technology has been widely used in urban security,unmanned vehicles,medical health,military reconnaissance and other fields.Deep learning-driven computer vision technology has the advantages of accuracy,flexibility and robustness,and strong ability to resist disturbances such as light changes,visual occlusion,size changes,and the proximity of targets and backgrounds that easily occur in aerial photography scenes.Among them,small target detection is a research hotspot in the field of vision.Since small targets have fewer available pixels compared with conventional targets,lack sufficient appearance representation,and are difficult to distinguish between targets and intricate backgrounds,aerial image small target detection achieves localization and recognition of a large number of small targets with high detection difficulty,and has high research value and wide application value.To address the above problems,this paper conducts research on aerial image ground small target detection method based on deep learning to achieve target detection architecture and achieve better performance.The details of the research are as follows:1)In this paper,we improve the channel attention mechanism and design a multi-scale attention component,which avoids the dimensionality reduction of the traditional mechanism and facilitates the information transfer of the network between different channels.This module enables the network to focus on key features,suppresses background redundant information and noise interference to a certain extent,and expands the perceptual field of the model in combination with the spatial pyramid structure.2)In this paper,we design a lightweight network MS-ECADet without anchor frames,which improves the feature extraction network and activation function and reduces the memory access cost.MS-ECADet incorporates a multiscale attention component that makes full use of features in different dimensions to mitigate the drastic deformation and uneven distribution of aerial small targets.Dynamic label assignment and new loss functions are used in the training strategy to mitigate the imbalance of positive and negative samples from a global perspective.The speed and accuracy of the lightweight model are balanced.3)To test the effectiveness of the aerial photography small target detection algorithm in this paper,an aerial photography image small target detection system is built.Based on the WPF framework and ONNX library,MS-ECADet is encapsulated and the system architecture and interface are designed.The human-computer interaction operation of the system is introduced,and the application of the improved target detection algorithm in the encapsulated case is realized.