Research On Remote Sensing Image Object Detection Based On Deep Convolutional Neural Network

Remote sensing technology is a comprehensive technology based on electromagnetic wave optical theory using high-precision sensors to image the electromagnetic wave information radiated from long-range targets to explore and study various visible scenes on the target surface,which is widely used in various applications such as mapping,military reconnaissance,maritime rescue and route planning.Remote sensing image target detection techniques based on convolutional neural networks have become a popular research direction in the field of computer vision due to their efficiency,speed and robust robustness.Due to the special imaging method,the target detection of remote sensing images requires corresponding improvement measures compared with ordinary images.In this paper,the following researches are carried out on remote sensing image detection algorithms:(1)Aiming at the complex imaging background,small size,large number of targets and dense arrangement in optical remote sensing images,the target detection accuracy is not high,and it is easy to miss detection and false detection.An improved candidate region algorithm is proposed.Firstly,an attention mechanism module is reasonably embedded in the backbone network to emphasize the target information to suppress the background information;secondly,a feature fusion method is designed to fuse the feature information of each layer in the convolutional neural network to enhance the small target detection capability of the model.In addition,the Anchor is optimized through the information statistics of the dataset;finally,the loss function and the pooling function are improved.Compared with the original algorithm,the m AP value of the improved method is significantly improved by 11.5%.Compared with other mainstream algorithms,the detection accuracy is more dominant,and it has better performance in remote sensing target detection.robustness and generalization ability.(2)An improved Mask R-CNN rotating frame detection algorithm is proposed for remote sensing warship targets with large aspect ratios and arbitrary placement angles,resulting in detection frames containing redundant information and mutual occlusion.A multi-angle rotation factor suitable for warship detection is added to the candidate area network and the IOU calculation method is fine-tuned,so that the warship detection result is a rotation frame that is more suitable for the actual target.Due to the complex imaging background of remote sensing images and the difficulty of detection,using depth-separated deformable convolution instead of standard convolution and optimizing the feature fusion structure makes the model maintain the detection speed while making a certain improvement in detection accuracy.A series of experiments show that the remote sensing warship detection model of the algorithm in this paper has achieved a good balance between detection accuracy and detection speed,and has high practicability.