Research On Object Detection Algorithm Of Remote Sensing Image Based On Transfer Learning

With the development of remote sensing technology,remote sensing information has been gradually introduced into various application scenarios as an auxiliary.Object detection is the key technique in the analysis of remote sensing information.Different from the classification problem,the specific category and location information of the ground object should be determined in the process of detection,so it is more challenging.In this paper,the aircraft detection task is taken as an example of object detection in high resolution remote sensing image.And the object detection task is studied in combination with transfer learning.The specific research content includes the following aspects:First of all,the basic theory of transfer learning and related methodology are studied.At present,the similarity between domains is the core of transfer learning,therefore,this section focuses on seven criteria used to measure the similarity between domains.In addition,in order to study the deep object detection algorithm,four deep transfer learning methods are studied first as the theoretical basis of the follow-up research.Secondly,Faster R-CNN algorithm in natural image is applied to the task of remote sensing image aircraft object detection.In order to solve the problem of small sample of remote sensing image,model-based transfer learning is applied.In this paper,two kinds of training methods,approximate joint training and non-approximate joint training,are compared.In addition,comparative experiments are set up to analyze the impact of the structure of transfer network and the number of samples on the results.The results show that the more complex the network,the higher the detection accuracy.The results show that when the approximate joint training method is applied,the more complex the network,the higher the detection accuracy.In addition,when ZF network is selected as the source model,the larger the number of training samples,the better the effect of model transfer.The detection accuracy of the model trained on a small-scale dataset consisting of about 1000 images is 81.41%.In the condition of scarce samples,the detection accuracy can meet the requirements of general detection tasks,which proves the effectiveness of transfer learning.Finally,Domain Adaptive Faster R-CNN(DA Faster R-CNN)object detection algorithm is proposed,which is based on the domain adaptation in transfer learning.DA Faster R-CNN object detection algorithm adds two domain adaptive structures based on Faster R-CNN.The improved algorithm can be used in the situation that the training data and test data are in different distribution,therefore,the object detection task can be successfully completed in low quality images.The object detection experiment based on low brightness aircraft datasets proves the effectiveness of the improved algorithm.Compared with the original Faster R-CNN object detection algorithm,the detection average precision of DA Faster R-CNN algorithm is greatly improved.