Remote Sensing Target Detection Based On Deep Feature And Attention Mechanism

With the continuous breakthrough of my country's aerospace science and technology,different types of remote sensing satellites have entered predetermined orbits,and my country's remote sensing technology has officially entered a period of vigorous development.Target detection has a huge impact on the military and civilian fields in the intelligent interpretation of synthetic aperture radar(SAR)images.Compared with optical images,low resolution,large speckle noise,and uneven target distribution all lead to poor synthetic aperture radar imaging effects and increase the difficulty of target detection methods.Constant false alarm rate detection is one of the general algorithms for SAR image target detection,but the different samples and clutter models will cause the detection accuracy of this algorithm to decrease.In addition,most of the saliency algorithms for different scenarios require steps such as sea and land segmentation to build models,and the performance decreases significantly.The method based on deep learning adopts an end-toend learning method,and has good detection results for SAR images of far seas,near shores and complex backgrounds.Based on this,this paper proposes a novel target detection model.In the feature extraction module,a multi-path aggregation deep feature pyramid network is proposed to extract the semantic information and location-spatial relationship features of the image,and suppress the clutter information interference and redundancy of the image.For information,in the detection part,a cascaded target detector based on attention guidance is used.Although the model parameters are increased,the ability to express the characteristics of the network is improved.The main contributions of this article are as follows:(1)Data augmentation.The data set used in this article is relatively small,with only1160 images.During the training process,this article oversamples the images containing small targets.Secondly,the small target size is within ±20%,and the rotation is zoomed within ±15°.Copy and paste in the image.Experiments show that this method can effectively improve the diversity of small target positions and balance the performance of the detector in the detection of small targets and large targets.(2)Deep feature pyramid network.Optical images often use color channel histograms to extract image color texture features.The SAR image not only has the attitude change in the horizontal size,but also contains the spatial structure information.The deep feature pyramid network of multi-path aggregation in this paper fuses the semantic information and spatial structure information of the input image.The interference of clutter factors that affect the accuracy of small target detection is removed.(3)Attention-guided cascade detector.Following the idea of attention mechanism and the idea of hollow convolution,this article introduces a multi-scale attention cascade module,and then through the convolution block and feature map cascade operation,the high-level feature map retains the low-level feature map information and reuses The main function of the first few layers of feature maps is to enhance the sparsity of the network,reduce the interdependence of parameters in the network,and alleviate the problem of overfitting.After the improvement of the above three points,this paper has been verified by comparative experiments.The accuracy and recall rate of the algorithm in this paper have good detection results,and the end-to-end training method has a wider application range.