Multi-type Object Detection For Large-field Remote Sensing Images Based On Deep Visual Perception Modeling

With the diversification of remote sensing satellite platforms and the improvement of spatial,time and spectral resolution,detection technology of space remote sensing remote sensing has made continuous breakthroughs.The object detection technology based on large-field remote sensing images has gradually applied urgently in the military and civilian fields,especially in the era of rapid development of artificial intelligence.However,the complexity and diversity of remote sensing scenes,the change of illumination and different weather conditions result in the instability of image quality,the significant difference of objects in the scene,and the large difference of scales between objects.At present,the mainstream remote sensing image objects detection method still has some problems to be solved urgently.Based on the current deep learning related theories and aiming at the related difficulties in the large-field of remote sensing object detection,this paper proposes a Multi-type object detection for large-field remote sensing images based on deep visual perception model.(1)Aiming at the difficulty of complex remote sensing scenes in a large-field and many false alarm interference factors,this paper proposes to construct semantic-visual features in the object detection network to represent the features of the object scene.This method mainly uses the unique connection mode between layers in the densely connected convolution module to extract rich semantic features.Based on the visual attention enhancement module extracts visual features in the channel and spatial dimensions.The feature information of the two dimensions is fused to construct deep structural features.Through a series of comparative analysis experiments,this method is effective in identifying remote sensing objects.(2)Aiming at the difficulty of large scale differences,small proportion and dense arrangement objects in the same scene,this paper proposes a method of multi-scale feature fusion in the object prediction stage.By further integrating semantic-visual features,the ability to represent object features is improved.In addition,feature maps of different scales are fused to construct a multi-level receptive field region,thereby overcoming the scale difference characteristics of the object in the scene.This algorithm not only achieves effective prediction for multi-scale objects in the object prediction stage,but also achieves reliable detection for small and dense arrangement objects.(3)Aiming at detecting false alarms in the detection results,this paper focuses on the analysis of false alarms and missed detection objects,and proposes a false alarm removal method based on lightweight convolutional neural networks.This method reduces the depth of the network on the basis of identifying the network(Peleenet).And use deep separable convolution to strengthen the connection between network layers.Therefore,the proposed method is a more suitable for false alarm removal from remote sensing objects.In addition,in the stage of false alarm elimination,this paper also proposes a differential confidence range extraction strategy,which effectively avoids a large number of false alarm discrimination,and further improves the speed of false alarm elimination.Experimental results show that this method has remarkable performance in eliminating false alarms.On the basis of the above-mentioned key technologies,this paper designs a software for large-field and multi-type remote sensing object detection to achieve accurate detection of objects in various types of complex backgrounds.In this paper,supported by a variety of public standard databases with complex large-field of remote sensing data,the effectiveness of the algorithm is verified through quantitative verification and comparative analysis experiment.