Fine-grained Classification And Detection Algorithms For Aircraft Targets Incorporating Broad Learning System

The refinement of high-resolution Earth observation systems and the upgrading of remote sensing platforms have promoted the research application of sub-metre spatial resolution remote sensing images in the detection and classification of fine-grained aircraft.There are problems involving numerous parameters,complex computations and great weight file size,an imbalance between the accuracy and efficiency of lightweight algorithms,and poor applicability for mobile deployment in fine-grained aircraft classification and detection algorithms currently.The aim of this article was to propose an improved algorithm for fine-grained detection of remote sensing aircraft targets,lightweight aircraft target detection,and fusion broad learning system for aircraft fine-grained classification by integrating multiple network modules and mechanisms based on the YOLOv5 object detection algorithm.The performance of the algorithm was tested and evaluated using the FAIR1M ten class aircraft target finegrained dataset with the following details and results.1.To address the issue of low precision in fine-grained detection of remote sensing aircraft targets,we first introduced a circular smoothing label-optimized angle classification algorithm in the YOLOv5s algorithm for aircraft target rotation detection to reduce the overlap of detection frames and background impact;then added a recursive gated convolution module to the YOLOv5s algorithm backbone network to enable the model to achieve high-order spatial interaction;and finally used a parameter-free attention mechanism to improve the feature representation ability of the model in a lighter way.The model testing results on the dataset in this article showed that the improved algorithm improved the mean average precision of the aircraft by 2.4%,and the number of model parameters,weight file size,and amount of computation increased by 1.8%,1.8%,and 0.6%,respectively.The inference speed on the server GPU could still reach more than 30 FPS,improving the model’s ability to detect fine-grained aircraft targets while also meeting the demand for real-time detection of fine-grained aircraft targets.2.To solve the problem of unbalanced detection accuracy and efficiency of lightweight algorithms for remote sensing aircraft target detection,firstly,the YOLOv5s algorithm was combined with the ShuffleNetv2 network to achieve algorithm lightweight;secondly,the ESNet module,Stem block,and simple pyramid pooling module were applied to enhance the backbone network feature extraction and feature representation capability of the algorithm model,while keeping the model lightweight;and finally,we also used the same parameter-free attention mechanism and introduced the SIOU loss function to improve the convergence efficiency of the algorithm.The model inference results showed that compared with the YOLOv5s model,the number of model parameters,amount of computation,and weight file size decreased by 78.74%,74.49%,and 85.55%,respectively.The inference speed on the mobile CPU and GPU was improved by 71.7%and 64.0%,respectively.And compared with the YOLOv5s-ShuffleNetv2 model,the detection accuracy rised to 98.59%with an improvement of 0.44%.The improved remote sensing aircraft target detection model was lighter,more efficient,and easier to deploy.3.To deal with the problem of low fine-grained classification accuracy of remote sensing aircraft model,a method combining broad learning image classification network with YOLOv5 classification algorithm was proposed to improve the fine-grained classification ability of the model.Firstly,a broad learning hybrid stack model was constructed using the broad learning system and its variants as well as the stack structure.Then,the aircraft model features extracted from the YOLOv5 classification algorithm network were used for training,and the precision of fine-grained classification was improved by enhancing the ability of feature extraction and nonlinear expression of the algorithm.Compared with the YOLOv5 classification algorithm,the number of parameters and weight file size of the improved model increased by 0.01M and 0.09MB,respectively,and the accuracy of aircraft type classification improved by 0.8%.The test set(4616 images)increased inference time on mobile device CPU by only 0.3775 seconds,which improved the accuracy of fine-grained aircraft classification without affecting the efficiency of model deployment on mobile devices.The algorithm improvement scheme proposed in this article effectively enhanced the fine-grained detection ability of model aircraft targets,improved the deployment ability and detection efficiency of the mobile end of the model,increased the fine-grained classification accuracy of model aircraft,and provided new solutions to the problems existing in remote sensing aircraft fine-grained classification detection tasks.