Research On Remote Sensing Image Target Detection Algorithm Based On YOLOv4

With the continuous progress of hardware technology,remote sensing technology has also been greatly developed,and people can obtain remote sensing images with higher quality.Remote sensing images have more important significance in both civil and military applications,so how to better carry out target detection tasks in remote sensing images is a problem that cannot be ignored.In remote sensing images,there are many problems,including many small target objects in images,different scales of target objects in images,more complex backgrounds in images and uneven distribution of target objects in images.These problems bring difficulties to the task of target detection in the field of remote sensing images,therefore,how to quickly locate targets and accurately detect target object types in remote sensing images has become a difficult problem to be solved.Since the traditional target detection means rely on manual extraction of features,there are defects in detection speed and accuracy.In recent years,with the progress of computer technology,deep learning-based target detection technology has replaced the traditional target detection means with its powerful computational performance.However,due to the difficulties in remote sensing images,it is still a problem that needs to be solved for the current target detection technology.In this paper,after analyzing the target characteristics in remote sensing images,the target detection method based on deep learning is improved as follows:(1)For the problem that the size of remote sensing image itself is too large,a multi-scale sliding window cropping method is designed by analyzing and comparing the rule grid cropping method with the sliding window cropping method.The method first counts the types,numbers and related sizes of the target objects.After that,different crop sizes are set according to different kinds of objects.And the dataset is enhanced using the Mosaic data enhancement method when the dataset is input to the model for training.(2)For the problem of small target loss in feature extraction network,a DN-NS module is designed based on the idea of attention mechanism.The module consists of a densely connected part and an attention part,which can enhance the expression of features while retaining more information about small targets.(3)For the problem of uneven distribution of target features at each scale in the feature fusion network,a feature fusion pyramid structure of B-PAN is proposed based on the idea of Bi FPN bidirectional feature fusion pyramid.This structure enriches the amount of feature information in the output feature map through the B-PAN structure by setting a top-down feature enhancement path,passing the features of small targets downward to fusion,and enhancing the reuse of feature information in the middle layer by jumping connections.Compared with several common target detection methods,the method in this paper is innovative in the above three aspects.After comparison experiments on DOTA dataset,the improved model in this paper can have better performance,which shows the effectiveness of the improved model in this paper.