Multi-type Object Identification For Remote Sensing Images Based On Deep Information Fusion

With the diversification of remote sensing satellite platforms and the improvement of imaging resolution,remote sensing images have been widely used.At present,there is an urgent application demand for remote sensing image intelligent processing technology in both military and civil fields.In recent years,with the rapid development and promotion of deep learning technology,the application of intelligent image processing in remote sensing field based on deep learning technology is gradually rising.However,because the large field of optical remote sensing image is a top view image,and there are many problems such as complex scene,many false alarm interference factors,large scale difference between targets and dense berthing,the mainstream natural scene depth learning algorithm is often difficult to achieve practical application in the field of remote sensing.Therefore,this paper proposes a multi-type object identification method based on depth information fusion.The object identification technology includes two key technologies: object detection and object classification.（1）In the detection stage of optical remote sensing image,a oriented object detection method based on depth information fusion is proposed.Firstly,the embedded attention module and multi-scale feature fusion detection method are used,in order to extract the image features effectively.At the same time,this method also can solve the problem of large scale difference between object in optical remote sensing image.Secondly,by predicting the angle of the object to achieve the oriented object detection method,this regression method can effectively solve the problem of dense berthing,and more conform to the requirements of remote sensing image multi-type object detection.Meanwhile,in order to solve the problem that there are too many set up anchors in the general oriented object detection method,this paper adopts the adaptive training sample selection.We also proposed a cascade module to improve the detection accuracy without increasing the amount of calculation.Finally,the loss based on gradient density is applied to the method.This loss can solve the problem of imbalance between positive and negative samples.In addition,it also has the effect of hard negative mining to further improve the network performance.（2）In the classification stage of optical remote sensing image,a light-weight network based on asymmetric convolution is proposed to process the above detection results,further eliminate false alarms and achieve fine classification.This paper first selects an excellent lightweight network Efficient Net as the baseline network,and adds dropblock to alleviate the problem of over fitting during training.In addition,in order to improve the classification accuracy,this study promote the robustness of the rotating object by changing the classical convolution to the asymmetric convolution.On the basis of the above method,a object identification software for large field of optical remote sensing image is implemented,which achieves high identification accuracy for multi-type object.Supported by the large open optical remote sensing data set DOTA,this paper verifies the effectiveness of the method through qualitative and quantitative experiments.