Research On Remote Sensing Image Object Detection Based On Context Information And Deployment Optimization

With the rapid development of optical remote sensing technology,the application fields of remote sensing image object detection technology are becoming more and more extensive.Remote sensing images have the characteristics of complex background information,dense object arrangement,and large-scale changes.The existing object detection algorithms are insufficient in the application of context information.In this thesis,new object detection algorithms are proposed to effectively use context information to improve detection accuracy and multi-scale detection performance.The main research works are as follows:Aiming at the problem of insufficient matching of candidate objects with large-scale changes caused by the hard discrimination method of object proposal determination in the object detection algorithm based on anchor.Designed an object proposal generation algorithm based on context information,adaptively generate more semantically representative object proposals with a high matching degree through multi-channel information association,combine the generated scale information and switchable dilated convolution to dynamically select the most appropriate feature receptive field.Aiming at the problem that a single intersection over union indicator cannot fully evaluate the matching degree between the object proposal and the ground truth,a multi-indicator comprehensive matching measurement model is designed,and a positive and negative sample assign probability discrimination algorithm is established,which can effectively filter candidate areas with low matching degrees,reduce false alarms.Aiming at the problem of inconsistent optimization objects when sharing object features in positioning and recognition tasks in anchor free object detection algorithms,a multi-task information-enhanced object detection algorithm is designed,which uses object prior information and multi-scale semantic segmentation to enhance positioning and recognition tasks respectively local context information,alleviating information bias problems.Aiming at the problem that the algorithm does not effectively use remote sensing scene information,a global context module is designed to extract global scene information for positioning and recognition.Experiments prove that the algorithm effectively improves the detection performance.Aiming at the problem of poor detection performance of existing object detection algorithms for multi-scale objects,a multi-scale feature fusion detection algorithm based on the MI-Transformer model is designed.The linear context function is used to extract the same-level context association,the cross-scale context encoding is used for multi-scale information interaction,and the weight decoding is used for feature fusion.Experiments prove that the algorithms improve the multi-scale detection performance of multiple algorithms.In response to actual engineering application requirements,a low-precision quantization algorithm based on information gain was designed,combining knowledge distillation technology based on mean square loss and Tensor RT inference acceleration library,the deployment optimization technology of the algorithm model was studied,and the optimized model deployment on the NVIDIA Xavier embedded platform was realized.While ensuring accuracy,the scale of the optimization model is reduced by more than 60%,and the inference speed is increased by more than 2 times.