Hyperspectral target detection method based on multi-instance spatial spectrum information joint extraction

Abstract

The invention provides a hyperspectral target detection method based on multi-instance spatial-spectral information joint extraction, which comprises the following steps: on the basis of a multi-instance framework, obtaining samples in balanced distribution by constructing positive and negative sample pairs, inputting the sample pairs into a spatial-spectral information joint extraction network, and constraining the network by using classification loss to obtain a hyperspectral target detection result; the network can be optimized towards a correct direction; a positive and negative sample pair is constructed by setting any number of pixels in each data packet, and the confidence degree of the pixel to a target is obtained through a pixel-by-pixel test. According to the method, the overfitting problem caused by the particularity of the hyperspectral data is effectively avoided, the target detection effect is better, the used network is an end-to-end network structure which directly classifies pixels, the method can adapt to the situation that a packet contains different sample numbers, and the detection efficiency is improved. Therefore, during testing, a single pixel can be directly input to obtain the confidence coefficient of the single pixel, so that the method has relatively high universality.

Description

technical field [0001] The invention relates to the technical field of image processing, in particular to a hyperspectral target detection method, which can be used for explosive detection and fine classification of crops. Background technique [0002] Due to the rich spatiotemporal information of hyperspectral images, it has been widely used in various fields such as explosive detection and fine classification of crops in recent years. However, due to the accuracy of the sensor, a pixel marked as an object in the hyperspectral image does not necessarily exist in the ground truth, but indicates that the object exists in a certain range of space including the pixel. In addition, because the background is complex and diverse, and the number of targets is much smaller than that of the background in most cases, it becomes difficult to detect targets in hyperspectral images. [0003] Multi-instance learning originated from drug activity detection. With its increasingly wide appl...

AI Technical Summary

Problems solved by technology

[0008]The above methods, combined with machine learning and deep learning technology, have improved performance compared to traditional methods, but these methods have higher requirements for data

However, hyperspectral data often has the problem of insufficient targets in the data, that is, the target to be detected rarely appears or even does not exist in the scene, which will lead to the problem of target imbalance in data allocation.

Using these data to train the model in the above method is prone to overfitting, resulting in a decline in detection performance

Images