The Application Of Recurrent Feedback Convolutional Neural Network For The Interpretation Of Remote Sensing Images

With the continuous progress of deep learning theory,intelligent interpretation of remote sensing images has gradually become a hot research topic in remote sensing earth observation field.Hyperspectral image(HSI)classification and synthetic aperture radar(SAR)image ship detection are two common applications in the fields of remote sensing,which play an indispensable role in civil and military fields.Due to the limited label samples in HSI classification,and the increase of network depth often causes the overfitting problems,at the same time,the pooling operation in the neural networks will also cause the loss of object detail information.In addition,SAR image data often contains more speckle noise,and small-sized ships and ship targets near the coast are prone to false alarms and missed detection.Aiming as the above problems,this article considers the characteristics of deep learning,and mainly focuses on the recurrent feedback convolutional neural network,forming two deep models for the HSI classification and SAR ship detection.The main research work is as follows:1.To solve the problems in HSI classification,recurrent feedback mechanism is introduced for the first time,combining attention mechanism,residual learning,and convolutional neural network to construct an unified deep framework(spatial attention-driven recurrent feedback convolutional neural network,SARFNN),including SARF2 DNN and SARF3 DNN models.In order to learn the spatial features in HSI data,an novel SARF2 DNN model is designed for the spatial feature extraction and classification of HSIs.Moreover,considering the geometric structure characteristic of the HSI data,the 3-D extended structure(i.e.,SARF3DNN)is further proposed to realize the spatial-spectral classification.In particular,with the help of the brain-like learning,the recurrent feedback structure can repair the lost detailed information caused by the pooling operation and the deeper structure.2.As for the ship detection in SAR images,we propose to use the neural network of core object recognition classification model(CORNet)as the backbone network,and by introducing the idea of recurrent feedback and attention mechanism,and a joint channel attention and recurrent feedback structure improved YOLOv3 model(channel attention-driven recurrent feedback neural network,CARFNN)is further built for SAR image ship detection.Finally,the effects of attention mechanism on multi-scale feature extraction is further studied in detail,and the experimental results verify the effectiveness and robustness of the proposed ship detection algorithm is verified.