Multi-strategy deep learning remote sensing image small target detection method

Abstract

A multi-strategy deep learning remote sensing image small target detection method comprises the following steps: marking a target in each remote sensing image, making a sample, copying the small target to other positions of the remote sensing image sample, and integrally forming a small target detection sample set; taking a deep learning target detection framework as a target detection model, training the target detection model by using a remote sensing image typical target detection sample set, then constructing a small target detection composite sample set, carry outing fine tuning training on the trained target detection model by using small target detection composite sample set, and carrying out training by increasing the training sampling frequency containing small target samples and increasing the number of 32-pixel scale anchor points in the fine tuning training process; and performing target detection on the remote sensing image by using the target detection model subjected to fine tuning training, establishing a multi-scale pyramid of the detected remote sensing image, increasing the number of 32-pixel scale anchor points, and reducing an IoU threshold value confirmed by a 32-pixel scale candidate region to 0.5. According to the invention, the accuracy of small target detection is improved.

Description

technical field [0001] The invention relates to a remote sensing image target detection and recognition technology. In particular, it involves a multi-strategy deep learning method for small target detection in remote sensing images. Background technique [0002] In the field of remote sensing image target detection, there are mainly the following methods for small target detection, but they all have some defects in accuracy and robustness: [0003] (1) The small target detection algorithm based on manual feature construction first looks for areas where there may be targets on the input original image, then extracts features from each area, and sends them to the classifier model for judgment, and finally the classifier model thinks that The area that is the target is subjected to post-processing operations such as screening to obtain the result. Due to the lack of effective image representations, people have no choice but to design complex feature representations and use v...

AI Technical Summary

Problems solved by technology

Due to the lack of effective image representations, people have no choice but to design complex feature representations and use various acceleration techniques to exhaust limited computing resources

[0004] (2) A small target detection method based on multi-scale deep learning. In a convolutional neural network, low-level features can often represent detailed information such as image textures and edges, while high-level features can often represent images well. Semantic information, but correspondingly, some detailed information will be ignored as the convolution pooling proceeds

Although this improves the resolution of the input image and is beneficial to the detection of small targets, it also brings other problems. The super-resolution model and the detection model are trained independently of each other. In the high-resolution input image generated by the super-resolution model It also includes unnecessary detection and unnecessary detection of objects and factors, and the increase in the resolution of the input image makes the overall architecture too heavy, the training and prediction time of the model will increase significantly, and the super-resolution will increase some targets that may be wrong details, reducing the likelihood that the practical application

Images