Full-convolutional-network aircraft detection method based on typical example mining

Abstract

The invention provides a full-convolutional-network aircraft detection method based on typical example mining. The method comprises the following steps: acquiring a training sample of a remote sensingimage; preprocessing the training samples to obtain a scale and a length-width ratio of an extended aircraft marking true value frame; carrying out automatic mining of a typical aircraft example; constructing a correspondence relationship between the typical aircraft example and the aircraft marking true value frame; carrying out full convolutional network training based on the typical aircraft example mining; and realizing aircraft detection by using the full convolutional network. According to the invention, the full convolution neural network is introduced into remote sensing image aircraft detection; clustering is carried out on an aircraft sample based on aircraft configuration features; the typical aircraft example is mined in training data; and thus a candidate target extraction network for different typical aircraft examples is trained. Therefore, performances of the candidate region extraction network are improved and thus the recall ratio of the aircraft detection is increased; and a false alarm suppression effect is realized.

Description

technical field [0001] The present invention relates to the technical field of digital image processing, more specifically relate to a kind of fully convolutional network aircraft detection method based on typical example mining. Background technique [0002] The artificially designed low-level features in the current aircraft detection method for remote sensing images are easily disturbed by the complex background in remote sensing images, and the description ability is relatively insufficient; while training multiple classifiers for different scenarios often leads to exponential growth in computation and time overhead , it is difficult to meet the actual application requirements of aircraft detection tasks. In addition, in the traditional aircraft detection method based on convolutional neural network, during the training process of the network, all aircraft samples used for training are considered to have high visual similarity, and are used as a unified whole to train th...

AI Technical Summary

Problems solved by technology

However, all aircraft samples participating in network training usually have relatively low visual similarity due to differences in models and sizes. Therefore, if all training samples are used to directly train the network without distinction, it is easy to reduce the performance of the training network and cause aircraft Detection recall drops

Images