Research On Small Object Detection Method Based On Feature Enhancement In Remote Sensing Images

Object detection is the basic and core task in the field of computer vision,which mainly refers to identifying the categories of objects of interest in the image and marking their position in the image with bounding boxes.With the development of deep learning technology,convolution neural network with unique advantages has greatly improved the performance of target detection algorithm,and has become the mainstream method to deal with the task of target detection.At the same time,with the continuous development of remote sensing technology,the number of high-resolution remote sensing images is increasing rapidly.Object detection based on remote sensing image has become a hot research topic,which has important applications in land use,disaster rescue,military reconnaissance and other fields.However,remote sensing image has the characteristics of long imaging distance,cluttered background,arbitrary object direction and so on,so the object detection task under remote sensing images has certain challenges.In view of the characteristics of remote sensing images and the shortcomings of existing methods,this paper studies the object detection method under remote sensing images based on the convolutional neural network.The main research work is as follows:Aiming at the problem that the cluttered background of high-resolution remote sensing images seriously interfere with the detection accuracy and the small target with insufficient information makes the detection difficult,a feature enhancement sub network is proposed in this paper.One of the two,the multiscale attention module models the global context information of the multiscale feature maps through attention structure.At the same time,the attention mechanism can guide the network to pay attention to the target area,suppress the interference of complex background and improve the discrimination ability of the network.Feedback connection is added in the feature enhancement pyramid module,which can feed back the multilevel feature maps generated by the feature pyramid to the backbone,and fuse the adjacent layer features from the bottom up during the feedback process.Through two feature extraction and fusion,the ability of feature representation is enhanced and the accuracy of small object detection is improved.In this paper,the ablation experiments and comparison experiments on HRSC2016 and DOTA datasets prove the effectiveness of the proposed method in remote sensing image target detection tasks.Aiming at the problem that the traditional regression loss treats all targets as the same and ignores the geometric characteristics of objects,an adaptive regression loss is proposed in this paper.The loss fully takes into account the geometric characteristics of the target in terms of area and aspect ratio,and can automatically adjust the loss value of the target according to the area and aspect ratio.By increasing the position error weight,width and height error weight of small objects and the angle loss weight of long objects,the prediction effect of the model on small objects and long objects is improved.In addition,the refined head is proposed in this paper.By fine-tuning the original detection head of the model,the detection effect can be further improved under the condition of reducing parameters and calculation amount.In this paper,the effectiveness of adaptive regression loss and refined head is verified by ablation experiments,and the comparative experiments show that the proposed method is superior to most existing remote sensing image objects detection algorithms.The results show that the proposed model improves4.29% on HRSC2016 dataset and 5.64% on DOTA dataset.