Lightweight model for On-Orbit optical object detection
As an important transport carrier and military target,aircraft detection in remote sensing images is important for aircraft rescue,early warning,and other fields.At present,the widely used neural network model has a complex structure and requires a large number of parameters,which limits the computing and storage resources of aircraft detection satellites.The efficiency and accuracy of satellite in-orbit detection need to be studied,and the computational structure must be optimized.Using neural networks in lightweight operation can reduce the computational costs and compress the overall framework.In this study,on the basis of a deep separable convolution neural network combined with deep separable convolution,the SwishBlcok bottleneck module was established by referring to the construction idea of a reverse residual structure.The characteristics of the network were simultaneously expanded in three aspects as follows:ResBlock_body was replaced with the overall design idea of the main framework of YOLO v4.Simultaneously,the channel attention mechanism of SENet was used for reference and integrated into the network structure.Different weights were given to the extracted feature maps and information.On the premise of maintaining channel separation,a separable convolution structure was used to improve the SPP structure and PANet structure;in this manner,both the number of model parameters and the memory dependence could be reduced.Moreover,the convolution layer and the batch normalization layer were merged to further accelerate forward reasoning.Drawing on the focal loss function,the loss function of object detection was improved to solve the imbalance between foreground and background data samples.The quality of algorithm restoration necessitates verification.In this study,objective evaluation indices were used to measure the algorithm from multiple angles.The public RSOD dataset and an internally produced dataset were used to compare the high-performance network models for algorithm verification.In terms of the rationality of the various improvements in the network model,verification experiments were conducted to measure the quality and processing speed of the algorithm.Then,the trained model was deployed on an embedded platform to verify the detection speed of the improved YOLO v4 algorithm model for on-orbit object recognition.The number of parameters of the proposed scheme was reduced by sevenfold compared with that of the original method,and the number of FLOPs was reduced by approximately 30 at a recognition accuracy of 94.09％.Subsequently,the experimental results were compared with the findings for the YOLO series,SSD,MobileNet,CenterNet,and other cutting-edge network models.The proposed algorithm outperformed the other methods.The proposed on-orbit object detection model can overcome the limitations of computing and storage resources,which traditionally cannot support high-precision complex models.The experimental results from ground and embedded platforms also prove that the proposed on-orbit object detection algorithm can effectively detect remote sensing targets based on detection performance.Future research may expand the scale of remote sensing datasets and improve the universality of model application scenarios
NETL
NSTL
万方数据
