Convolutional neural network-based multispectral image semantic cutting method

Abstract

The invention relates to a convolutional neural network-based multispectral image semantic cutting method. The method comprises the following steps of: respectively and independently convoluting eachdata channel of a multispectral image by utilizing a convolutional neural network; and fusing feature maps obtained after independent convolution of the data channels. According to the method, a network with multi-resolution input and multichannel independent convolution is utilized, so that the problem that standard U-NET networks only can accept one same-scale RGB-Gray image is effectively solved, the working efficiency of multispectral image semantic cutting is effectively improved and the image cutting precision is ensured.

Description

technical field [0001] The invention relates to a multispectral image semantic segmentation method based on a convolutional neural network. Background technique [0002] Currently, state-of-the-art semantic segmentation frameworks for RGB images commonly employ end-to-end deep convolutional neural networks (DCNNs). The current use of convolutional neural networks is to use some pre-trained models to classify objects. These commonly used models mainly include VGG, ResNet, etc. The DCNN for semantic segmentation often includes two parts, the first half is a commonly used DCNN network with better quality, and the second half is a network that maps feature maps to pixel labels. In order to save training samples, the pre-trained model parameters are directly used in the first half, and only the second half of the model parameters are fine-tuned. [0003] At present, the representative image semantic segmentation network is fully-convolutional network (FCN), and the initial vers...

AI Technical Summary

Problems solved by technology

If all low-resolution images are forcibly interpolated and enlarged to unify them with high-resolution images, some convolution operations on the low-resolution image parts will be invalid, not only losing a lot of computing time, but also possibly interfering with The result of image cutting

If the high-resolution image is down-sampled to make it consistent with the low-resolution image, a large amount of spatial information of the high-resolution image will be lost.

Images