Research On Aerial Target Detection And Recognition Technology Based On Machine Learning

With the rapid development of image acquisition and processing technology,Machine Vision is increasingly used in various fields,especially for image engineering of aerial vehicles.This paper mainly studies the hardware platform of image acquisition and the technology of image detection and recognition for aerial targets.This article mainly carries on the research from the following aspects:In view of the characteristics of detection and recognition of aerial targets,this paper proposes a target detection and recognition technology based on Machine Learning,especially Deep Learning.The basic principles of Deep Learning,the convolutional neural networks commonly used and the model of Deep Learning studied in this paper are introduced respectively.According to the algorithm characteristics of Deep Learning model,an image acquisition system based on the platform NVIDIA Jetson TX2 is built in this paper,focusing on the selection basis of hardware platform and interface conversion.For pixel loss in the detection of small targets in the far space,an S-Faster R-CNN(SmallFaster Region-CNN)model using ROI Align layer is proposed in this paper.The Faster R-CNN model has difficulty in detection of small targets due to multiple quantization of images in the ROI Pooling layer.On the basis of this,the ROI Align layer is proposed to replace the ROI Pooling layer for image size immobilization.The improved model improved the detection rate by 11.78%.Aiming at the fine-grained recognition in the detection of large near-Earth targets,a Bilayer Faster R-CNN with Feedback is proposed.The recognition of large targets is essentially a problem of fine-grained target recognition.In this paper,an idea of offline hard example mining is proposed.Feedback module and retraining module are added to Faster R-CNN model.The improved model improves the recognition rate by 3.81%.The theoretical analysis and experimental results show that the image acquisition and processing system designed in this paper perform well to realize high-speed acquisition of highpixel images,the accuracy of image processing system for small target detection in far space and fine-grained recognition of large target in near ground has been improved.