A remote sensing image target detection method based on deep neural network

Abstract

A remote sensing image target detection method based on a deep neural network, comprising: obtaining training samples and performing a scaling preprocessing operation on the training samples; Perform various types of labeling operations on the sample image; perform pre-training and fine-tuning operations on the selected deep convolutional neural network model based on the maximum number of iterations to obtain a fine-tuned deep convolutional neural network model; The fully connected layer of the network model is processed to obtain a fine-tuned deep convolutional network model; the remote sensing image target is detected based on the fine-tuned deep convolutional neural network model and the fine-tuned deep convolutional network model. This method has the advantages of fast detection speed and high detection accuracy, and can be widely used in target detection, target tracking, intelligent navigation, urban planning and other fields.

Description

technical field [0001] The technical field of the present invention is digital image processing and pattern recognition, involving image processing, deep learning algorithm, image classification, target detection and other content. Background technique [0002] Target detection in remote sensing images refers to detecting one or several types of targets (such as airplanes, bridges and houses, etc.) in large-scale high-resolution remote sensing images and marking their positions. Early object detection in remote sensing images was mainly based on template matching and shape priors. With the development of deep learning methods such as convolutional neural networks (CNNs) and their successful application in image classification, object detection, and other fields, deep The learned method has also been transferred to object detection in remote sensing images. [0003] At present, the main idea of ​​using deep learning algorithm to solve the problem of remote sensing image targ...

AI Technical Summary

Problems solved by technology

[0004] The existing technology mainly has the following problems: First, when using the region search algorithm to select candidate regions, a large number of candidate regions that do not contain remote sensing targets will be generated, thereby increasing the false detection rate of the detection algorithm. The workload of CNN model classification; second, the feature extraction ability of the shallow CNN model is insufficient, and the classification effect is not ideal, but because there are not enough remote sensing image samples, the fine-tuning of the deep CNN model cannot be completed, and the candidate area cannot be completed using the deep CNN model Third, the non-maximum suppression algorithm can remove overlapping detection frames, but cannot adjust the position of the detection frames, which makes the remote sensing image target detection results inaccurate

Images