Object Detection For Optical Remote Sensing Images Based On Multi-scale Deconvolutional Feature Fusion Network

Nowadays,object detection has been one of the most important tasks in computer vision field.However,deep convolutional neural networks are often used to detect objects for natural images,and they are rarely used for optical remote sensing images.Therefore,it makes sense for us to explore more effective object detection algorithms for optical remote sensing images based on deep convolutional neural networks.Based on the analysis stated above,this thesis uses deep convolutional neural networks to obtain and fuse the feature information of shallow and deep layers for object classification and bounding box regression.In terms of the main work in this thesis,it can be listed as three aspects:1.A method of object detection for optical remote sensing images based on the deconvolutional context information fusion network is proposed.Firstly,the objects in optical remote sensing images are located to get the image cutting blocks with random offsets.Meanwhile,in order to match the number of data and the number of parameters in the network to avoid overfitting,the strategy of data augmentation is applied to the data generated by cutting operation to form the training dataset.Then,a deconvolutional module is added to connect the context information in the original SSD network and the new network is named as deconvolutional context information fusion network.Finally,the deconvolutional context information fusion network is used to classify and regress objects.The key of this method lies in the introduction of the deconvolutional module,overcoming the problem of the low precision ratio of optical remote sensing images.As to the effectiveness,it is proved by experiments.2.A method of object detection for optical remote sensing images based on the refined deconvolutional shallow feature fusion network is proposed.This network includes an anchor refinement module and an object detection module.Firstly,the anchor refinement module is used to filter anchors,and then the retained anchors are sent to the object detection module for classification and regression.Moreover,a feature fusion block is led in,and its purpose is to fuse the shallow features extracted by the channel.The key of this method lies in the extraction and fusion of shallow features,improving the recall ratio for object detection.Compared with other methods,its effectiveness is proved.3.A method of object detection for optical remote sensing images based on curvelet refined deconvolutional shallow feature fusion network is proposed.Firstly,the curvelet transform is used to extract the edge information from many directions of objects,what is more,the edge information extracted and original images are fused to get abundant feature information.Then,they are sent to the refined deconvolutional shallow feature fusion network.The key of this method is to obtain the edge information of objects by curvelet transform,enhancing the representation ability of the network.Compared with other methods,the performance of this method is better,which is proved by experiments.