Hyperspectral image classification method based on self-paced learning double-flow multi-scale dense connection network

Abstract

The invention provides a hyperspectral image classification method based on a self-paced learning double-flow multi-scale dense connection network, and the method comprises the following steps: 1, reading a hyperspectral image for preprocessing, including unsupervised waveband selection dimension reduction and image partitioning, and dividing a training set and a test set; 2, establishing a local spatial spectrum feature extraction branch and a global spatial feature extraction branch, adaptively fusing the features of the two branches, and finally connecting full connection and a Softmax layer; 3, respectively inputting the local image block and the global image block corresponding to the same pixel point into two branches, training an effective hyperspectral image classification model by taking a minimized cross entropy loss function as a target, and selecting training data by utilizing a self-paced learning framework; and 4, after the training is finished, inputting the test set data into the network to obtain a prediction category output by the model.

Description

technical field [0001] The invention relates to the technical field of hyperspectral image processing methods and applications, in particular to a hyperspectral image classification method based on self-learning dual-stream multi-scale densely connected networks. Background technique [0002] Hyperspectral images have the characteristics of multiple spectral bands and continuous imaging in specified bands, which can capture rich object features. The rich information of hyperspectral images helps to classify and classify ground objects, biological tissues, and chemical materials. Therefore, hyperspectral technology has been widely used in many fields such as geological science, aerospace, and medical diagnosis. [0003] As an effective feature extraction method, deep learning has achieved excellent performance in many computer vision tasks such as image classification and segmentation, dense object detection, etc. Compared with traditional hyperspectral image classification ...

AI Technical Summary

Problems solved by technology

[0004] Most of the existing pixel classification algorithms for hyperspectral images are simple improvements based on ordinary classification algorithms, which do not fully consider the information redundancy and data characteristics of hyperspectral images, so these methods will have some shortcomings in actual classification tasks : 1. In order to effectively extract the spatial features of pixels, the hyperspectral image is divided into image blocks of different spatial sizes during preprocessing, and the global spatial features and local spatial features of the central pixel are extracted by using large-size and small-size image blocks.

However, this preprocessing process takes a long time and consumes a lot of storage resources

2. Due to the high spectral dimension of hyperspectral images and the limitation of GPU computing resources, the spatial resolution of the cut image blocks should not be too large, otherwise the amount of sample data is too large and difficult to train

Therefore, the spectral dimension of the image block often far exceeds the spatial dimension, resulting in redundant spectral information, which makes it impossible to efficiently extract spatial features

3. The traditional multi-scale feature fusion structure often uses multiple branches to extract features from image blocks with convolution kernels of different sizes, and then directly stitches the features of different branches together, but there is a lack of quality between these feature extraction branches. Contact, directly spliced ​​features cannot well represent the local features and global features of the image block

4. For high-resolution hyperspectral images, if the method of point-by-point pixel classification is used, the extracted training data will be huge. If the traditional method is used to train in random order, it will be time-consuming, difficult to converge, and the effect cannot be achieved. good

5. Traditional feature fusion adopts the method of direct feature splicing, but for the fusion of spectral features and spatial features of hyperspectral images, the two parts of features have different data distribution characteristics, so an adaptive fusion method is required

Images