Enhancing Aircraft Object Detection in Complex Airport Scenes Using Deep Transfer Learning
Zhong Dan,Li Tiehu,Li Cheng
DOI: https://doi.org/10.3788/gzxb20245304.0415002
IF: 0.6
2024-01-01
ACTA PHOTONICA SINICA
Abstract:Within the civil aviation airports of China,intricate traffic scenarios and a substantial flow of traffic are pervasive.Conventional monitoring methodologies,including tower observations and scene reports,manifest vulnerability to potential errors and omissions.Aircraft object detection at airport scenes remains a challenging task in the field of computer vision,particularly in complex environmental conditions.The issues of severe aircraft object occlusion,the dynamic nature of airport environments and the variability in object sizes pose difficulties for accurate object detection tasks.In response to these challenges,we propose an enhanced deep learning model for aircraft object detection at airport scenes.Given the practical constraints of limited hardware computational power at civil aviation airports,the proposed method adopts the ResNet-50 model as the foundational backbone network.After pre-training on publicly available datasets,transfer learning techniques are employed for fine-tuning within the specific target domain of airport scenes.Deep transfer learning methods are utilized to enhance the feature extraction capabilities of the model,ensuring better adaptation to the limited aircraft dataset in airport scenarios.Additionally,we incorporate an adjustment module,consisting of two convolution layers,into the backbone network with a residual structure.The adjustment module can increase the receptive field of deep feature maps and improve the model's robustness.Moreover,the proposed method introduces the Feature Pyramid Network,establishing lateral connections across various stages of ResNet-50 and top-down connections.FPN generates and extracts feature information from multiple scales,facilitating the fusion of features in the feature maps.This enhances the accuracy of multi-scale target detection in the task of object detection.Furthermore,optimizations have been implemented on the detection head,composed of parallel classification and regression branches.This detection head aims to strike a balance between the accuracy and real-time performance of target detection,facilitating the fast and accurate generation of bounding boxes and classification outcomes in the model's output.The loss function incorporates weighted target classification loss and localization loss,with GIoU loss used to calculate the localization loss.Moreover,we construct a comprehensive airport scene dataset named Aeroplane,to evaluate the effectiveness of our proposed model.This dataset encompasses real images of diverse aircraft in various backgrounds and scenes,including challenging weather conditions such as rain,fog,and dust,as well as different times of day like noon,dusk,and night.Most of the color images are captured from the camera equipment deployed in various locations,including terminal buildings,control towers,ground sentry posts and other places of a civil aviation airport surveillance system in China.The diversity of the dataset contributes to enhancing the generalization performance of the model.The Aeroplane dataset is structured adhering to standards and is scalable for future expansion.And we conduct experiments on the Aeroplane dataset.Experimental results demonstrate that our proposed model outperforms classic approaches such as RetinaNet,Inception-V3+FPN,and ResNet-34+FPN.Compared to the baseline method,ResNet-50+FPN,our model achieves a 4.9%improvement in average precision for single-target aircraft detection,a 4.0%improvement for overlapped aircraft detection,and a 4.4%improvement for small target aircraft detection on the Aeroplane dataset.The overall average precision is improved by 2.2%.Through experimental validation,our proposed model has demonstrated significant performance improvement in aircraft target detection within airport scenarios.The presented model exhibits robust scene adaptability in various airport environments,including non-occlusion,occlusion,and complex scenes such as nighttime and foggy weather.This validates its practicality in real-world airport settings.The balanced design of real-time performance and accuracy in our approach renders it feasible for practical applications,providing a reliable aircraft target detection solution for airport surveillance systems and offering valuable insights for the task of object detection.