Typical Target Detection From Large Scene Remote Sensing Image Based On Deep Learning

As a large-scale monitoring means,remote sensing plays an increasingly prominent role in national security and military fields.It is of great strategic significance to detect and recognize various typical targets from remote sensing images automatically.In recent years,deep learning has been widely applied in remote sensing image target detection and made outstanding achievements,because of its advantage of learning the hierarchical features independently.Nevertheless,the existing deep learning-based remote sensing target detection methods still have some deficiencies.Firstly,they mostly adopt typical feed-forward networks in feature extraction,such as VGG and ResNet,and only employ deep-level features for prediction which lacks the effective use of different levels of features,resulting in serious loss of small and dense targets.Secondly,these methods do not fully consider the imbalance of positive and negative samples that is more prominent due to the huge proportion disparity between the target and background in the remote sensing image.Additionally,most methods are still in the stage of detecting small images with single scenes,which fails to form a complete detection system that can be directly applied to large scene remote sensing images transmitted from the satellite platform.In view of the problems above,starting from two basic schools of deep learning-based target detection algorithms:one-stage detectors and two-stage detectors,this paper reaerches on the typical target detection of remote sensing image and focuses on the solution of large scene remote sensing image target detection.The main works and contributions can be summarized as follows:To solve the problem of weak response of deep features to small targets and dense targets of remote sensing images,a two-stage and multi-scale compact dense block-based detector:MSCDB Faster RCNN is proposed.By introducing dense connection and compressing dense blocks,this algorithm designs a compact feature extraction network called DenseNet-65,and innovatively combines it with feature pyramid network,which not only achieve multi-scale feature prediction but alse makes the feature expression of small targets more significant.The experimental results show that the proposed methd has made remarkable progress in detecting small and dense aircrafts,and cost less parameters and calculation consumption.Addtionally,the method has also been proved to obtain an excellent performance in the detection of five typical targets:aircraft,ship,large vehicle,small vehicle andoil tank.Aiming at the imbalance problem of positive and negative samples ignored by most onestage remote sensing target detection methods at present,an optimized RetinaNet algorithm called RS RetinaNet is proposed.Focal loss is used to alleviate the imbalance of positive and negative samples in the training stage.The balanced feature pyramid is proposed,allowing the network to employ the features around the target for auxiliary recognition.Deformable convolution with modulation is introduced which enables the network to extract features of target that possesses obvious geometry more accurately,such as ships and oil tanks.Balanced smooth L1 loss function is used to achieve a more balanced training between target classification and box regression by improving the gradient of easy samples in bounding box regression.The experimental results show that compared with the standard RetinaNet,RS RetinaNet greatly improves the performance.This paper presents a target detection method which can be applied to the large scene remote sensing image that is directly transmitted from satellite platform.Considering the largescale Characteristic of large scene remote sensing image,a method of cutting large scene remote sensing image into sub-images with overlapping is proposed,which not only avoids the problem of target distortion in the process of scaling,but also reduces the risk of large-sized target being truncated.The post non maximum suppression algorithm is proposed,which simplifies the detection process while solving the false alarm targets in the overlapping area of sub-images.Besides,the entire image-based and guidance-point-based target detection processes are both constructed.The latter can detect the designated area according to the priori position information,thereby greatly improving the detection efficiency.Based on the above work,a target detection software for large scene remote sensing images is implemented,which integrates functions of image preprocessing,guide point acquisition and preprocessing,target detection,detection results and state file output.The software can directly run without the environment of deep learning.