Object detection method for remote sensing image based on convolutional neural network based on pruning strategy

Abstract

The invention discloses a remote sensing image target detection method for constructing a convolutional neural network based on a pruning strategy. The method comprises the steps that first, a targetdetection training dataset is preprocessed; second, an initial network is pre-trained to obtain a dense network; third, for the dense network obtained in the second step, a dense-sparse training modebased on a network pruning technology is adopted to obtain a trained sparse network, and a dense training mode is adopted to train the trained sparse network to obtain a precise target classificationnetwork; fourth, the trained sparse network obtained in the third step is processed, and a sparse-sparse training mode is adopted to obtain a quick region extraction network; and fifth, a multi-scaleimage pyramid is generated from a target detection test dataset, then the quick region extraction network and the precise target classification network are sequentially used to perform two-stage prediction, and the position and a category tag of a target are obtained.

Description

technical field [0001] The invention relates to automatic construction of a convolutional neural network and target detection of remote sensing images. Background technique [0002] The spatial resolution of optical remote sensing sensors has been greatly improved in the past decade: the GeoEye, WorldView, and Pleiades series all have high spatial resolution; GSD, Ground Sample Distance) images. The advancement of sensor technology provides opportunities and challenges for image interpretation such as target detection. A large number of high-resolution images provide the possibility for deep learning to be applied to remote sensing image target detection, but large-scale data leads to excessive calculation of detection. . [0003] Relying on the advancement of Region Proposal technology and the development of neural network's ability to extract features hierarchically, most recent deep learning-based remote sensing image target detection methods have achieved good results....

AI Technical Summary

Problems solved by technology

[0004] The current remote sensing image target detection methods have certain limitations: (1) The area extraction techniques currently used in the field of remote sensing include Edge Boxes (window scoring method), Selective Search (grouping method), etc., using hand-craft features, Therefore, the data set cannot be effectively used to automatically learn deep features, resulting in inaccurate and slow extraction of candidate area blocks.

(2) Remote sensing images have the characteristics of large images and small targets. In the training phase of the network, the existing solution mode directly inputs the original image and its ground truth information into the network, which leads the network to learn more background information, less target feature information

Using a sparse network without retraining will make the network less accurate

Images