Research On Automatic Detection Technology Of Typical Targets Based On Visible Light Remote Sensing Images

With the rapid development of image sensing technology and the aerospace industry,remote sensing images have become easier to obtain and their resolution has been greatly improved.High resolution visible light remote sensing images contain a large amount of ground object information.So,it is particularly important to analyze and make full use of this information.Remote sensing targets detection is also a way to interpretate remote sensing images.The aim of remote sensing targets detection is to determine whether an aerial or satellite image contains one or more targets and to locate each predicted target in the image.In recent years,the detection of typical targets in remote sensing images has played an important role in many fields such as resource utilization,urban planning,military target identification and battlefield environment simulation.However,the large size of images,the complex environment background,and the differences in the characteristics of different targets all make the detection task extremely difficult.Therefore,the detection of typical targets in remote sensing images is still a challenging and meaningful research.This paper mainly focuses on the detection technology of two typical targets in visible remote sensing images,namely,marine ships and parked aircraft.Starting from two different technical implementations,traditional machine learning and deep learning,the research is carried out on some key technologies,such as saliency detection,image filtering,feature extraction and classification,and convolutional neural network model construction,etc.The main contents are summarized as follows:1.In the task of detecting ship targets on the ocean surface,the background of the sea surface is relatively simple and uniform.So,the visual saliency method is introduced to detect salient areas that are different from the overall background in the wide remote sensing image,which is used to quickly search for potential targets.However,due to the characteristics of the sea surface and the weather at sea,there are often clouds,fog,sea clutter,and ship sailing wakes in remote sensing images,which bring interference to the target detection.In order to suppress the interference and highlight the target,a method combining local saliency analysis in the spatial domain and global saliency analysis in the wavelet domain is proposed.The saliency analysis method can effectively suppress multiple interferences and highlight the salient areas of the ship target without any prior knowledge.At the same time,the potential targets are identified by designing a dual discrimination mechanism based on ship target shape feature parameters and feature classification.From the experimental results,this dualstage detection method can eliminate false alarms and screen out real targets.Finally,the Randon transform is introduced to analyze the primary direction of the detected ship targets and determine the target azimuth.2.The ground background is more complex and diverse.The saliency method is not suitable for extracting potential aircraft targets from ground background.According to the inherent structural characteristics of the aircraft target,an image filtering method based on the fourth-order circular harmonic function is designed to detect the center point of the potential aircraft target.Due to the aerodynamics,aircraft targets generally have a criss-cross structure.If the center of the target is taken as the center of a circle,the inherent structure of the aircraft will cause regular changes in the gray value signal on the circumference,which show a four-period characteristic of alternating light and dark.This waveform change can be identified by designing a convolution template for calculating the signal correlation to determine the target center position.In order to locate objects at different scales in the image,an image pyramid is introduced to perform multi-scale processing on the original image.At the same time,the Fourier HOG feature with rotation invariance is also deeply studied,and the method of local feature aggregation descriptor is introduced to re-encode the feature to reduce the feature dimension and improve the feature representation ability.The final experimental results show that the overall detection method has achieved good detection results on the public RSOD dataset.3.As deep learning-based methods are insensitive to the class of the detected target,a network model is designed for the task of detecting multi-class rotated targets in remote sensing images,especially dense ship targets moored near shore.Based on the horizontal bounding box detection model–YOLOv5,additional channel for target angle prediction is added to the network to achieve target rotation detection.To address the problem of loss mutation caused by the range of angle definition during model training,the regression problem of angle prediction is transformed into a classification problem,and a method of discrete encoding of angles with circular smooth labels is introduced.The activation function of the original model is also improved.And the BIFPN module is introduced to fuse multi-scale features in order to enhance the feature extraction capability of the network and thus improve the detection capability of the model.Finally,ablation experiments were conducted on the HRSC2016 dataset to illustrate the effectiveness of the relevant improvements.Comparative experiments are also conducted on the DOTA dataset to illustrate that the proposed rotation detection model can also achieve good results in the detection task for multi-category typical targets.In summary,this paper presents an in-depth study of a variety of key technologies involved in the task of automatic detection of typical targets in visible remote sensing imagery,and proposes some new design and improvement ideas.The relevant conclusions and results obtained in this paper have a certain promotion effect on the automatic interpretation system of remote sensing images.