High-resolution Optical Light Remote Sensing Image Target Extraction Based On Domain Information-assisted Deep Learning

With the rapid development of Earth observation and remote sensing imaging technology,the number of remote sensing satellites is increasing rapidly,and the image quality is greatly enhanced,providing a large number of high-resolution visible light remote sensing images with high resolution,rich detail and clear texture.Therefore,the design of accurate and efficient remote sensing image target automatic monitoring and extraction algorithm not only has important theoretical research significance,but also has extensive applications in port management,maritime rescue,sea area monitoring,urban planning,key military target detection and other civil and military fields.Value.However,the traditional automatic target extraction method based on remote sensing image is implemented step by step by multiple independent “task modules” to obtain the final result,which is difficult to meet the needs of practical applications in both processing performance and processing efficiency.At present,with the support of big data and powerful computing power,deep learning has become one of the most effective machine learning methods for solving problems such as image understanding.The high-resolution visible light remote sensing image extraction method based on deep learning has become a hotspot in remote sensing image interpretation research.The deep learning method based on multi-layer neural network is characterized by its excellent feature learning ability.A significant advantage over traditional methods.However,unlike natural close-up images,the targets in visible light remote sensing images tend to be diversified,oriented,and densely arranged.At the same time,the textures and shapes between the targets are easily interfered with each other,and it is easy to receive artificial features with similar features.The impact of the background.At the same time,deep learning relies heavily on a large number of clean labels.However,obtaining such mark data is time consuming and labor intensive.Especially for remote sensing image annotation,it is often necessary to interpret the expert's participation,which is costly.The lack of a large amount of full-supervised data severely restricts the in-depth application of deep learning technology in automatic target extraction of remote sensing images.In view of the above-mentioned problems to be solved,this paper automatically extracts the remote sensing image ship and building target in complex background,fully exploits the domain information of the remote sensing image itself,and studies the deep learning method of automatic target extraction of remote sensing image with the help of domain information,reducing the number of markers.Sample dependencies improve target extraction performance.Specifically,the main innovations of this paper are:(1)A multi-orientation,multi-scale ship detection method for rotating full convolutional networks is proposed.The existing target detection method based on deep convolutional network performs target positioning through the vertical rectangular frame.However,in the remote sensing image,the ship targets tend to be different and densely distributed side by side,and the rectangular frames are difficult to be densely arranged side by side.The ships are extracted one by one.Therefore,in view of the dense and oriented arrangement of ship targets in remote sensing images,a rotating bounding box is proposed to describe the target position and region,and a multi-scale feature fusion full convolution nerve is designed.The network predicts the rotating bounding box and region of the target.The method uses a full convolution network to directly return to the target's bounding box,orientation angle,and target area mask.Using the target area mask as an additional supervised information can guide the network to learn better target feature representation.At the same time,the new method utilizes the different and multi-scale properties of the receptive field of the convolutional network feature map,and fuses the multi-scale convolution features to enhance the detection capability of multi-scale targets of remote sensing images.The experimental results show that the proposed algorithm can extract dense and side-by-side distributed ship targets,which greatly improves the detection performance and positioning accuracy.(2)Aiming at the problem that it is difficult to obtain the pixel-level clean mark of remote sensing image to train deep neural network for building area extraction,it is proposed to use the Geographic Information System(GIS)footprint information as an auxiliary mark to train the building area.Extracted full convolutional neural network.Due to the registration error and incompleteness of GIS data and remote sensing images,the mark provided by GIS data is not very accurate and contains a lot of noise.For this reason,a full convolutional neural network model based on noise label transition probability is proposed.The results show that the full convolutional neural network model with transition probability has certain robustness to marker noise,and it can significantly reduce the dependence on the refined annotation of remote sensing images while achieving efficient extraction of architectural targets.