Remote Sensing Image Object Detection Based On Attention Mechanism And Few-Shot Incremental Learning

Object detection is fundamental branch of computer vision,which can provide more valuable information for more complex tasks such as object tracking and image description.Its specific requirement is to accurately locate the position of the object and judge its category in the image,which is also one of the most popular research directions for many years.In recent years,thanks to the rise of neural networks,advanced algorithms in the field of object detection have emerged one after another and have achieved great development.However,with the rapid development of society and the emergence of new things,how to quickly fit new data has become a new problem,and in many cases,it is difficult and expensive to collect data with a large number of labels,therefore,it is necessary to study the few-shot incremental object detection algorithm.Due to the huge difference in size,angle,and direction of object instances,remote sensing image data has a poorer generalization ability than natural scene data,and is more challenging.This paper studies the problem of few-shot incremental object detection on remote sensing datasets:1.A weight imprinting method based on composite class representation is proposed.The benchmark algorithm in this paper uses the cosine classifier to classify,and calculates the instance feature mean to imprint the classifier weight parameter.In this paper,based on the K-means method,multiple representation vectors are generated for each new class,and the mean cosine similarity between the Region of Interest(Ro I)feature and the representation vector is calculated for classification,which expands the decision boundary of the new class and improve the representation ability of few-shot data.Compared with the incremental Mask-Two-stage Fine-tuning Approach(i MTFA),mean b AP50 is improved by about 1.67%,and mean n AP50 is improved by about 1.15% under the 10-shot setting of the remote sensing datasets.On the COCO dataset,b AP is improved by about 9.59%,and n AP is improved by about 0.71%.2.A parallel classifier structure based on knowledge distillation is proposed.Two classifiers are used in the roi head of the model,one of which is called ”Base Classifier”,whose parameters are the frozen base class prediction weight,and the other is called ”Incremental Classifier”,whose parameters are initialized to the weight parameters imprinted in 1.KL divergence of the output of the two classifiers is computed as distillation loss for consistency constraint.Compared with the i MTFA,under 1/5/10-shot data setting,mean b AP50 is improved by about 3.09%,and mean n AP50 is improved by about 1.34% on the DIOR dataset.On the DOTA dataset,mean b AP50 is improved by about 5.89%,and mean n AP50 is improved by about 0.34%.On the COCO dataset,mean b AP is improved by about 11.16%,and mean n AP is improved by about 2.6%.3.An adaptive module based on graph attention networks is improved.In the fine-tuning process,the graph attention module is used to adaptively update the classifier weights and Ro I features to improve the generalization ability of few-shot remote sensing data,thereby improving the detection accuracy.Finally under 1/5/10-shot data setting,the mean AP50 is about 51.52% on the DIOR dataset and about 31.04% on the DOTA dataset,and the mean AP is about 29.39% on the COCO dataset.