Abstract: In the case of zero sample remote sensing image scene segmentation without any observation data, there is no response reference, which results in long segmentation time and low accuracy. Therefore, a zero sample urban remote sensing image scene segmentation algorithm based on convolutional neural network is proposed. PCA ( Principal Component Analysis) and K-SVD ( K-Singular Value Decomposition) are used to denoise remote sensing images to suppress the patch effect. The denoised image is input into the Retinex enhancement algorithm to further improve the enhancement effect of zero sample urban remote sensing image. The mean shift algorithm is used to segment the remote sensing image scene to obtain the relationship between its pixels, and the convolution neural network is used to complete the accurate segmentation image scene. The experimental results show that the algorithm has high accuracy, high recall, high F-score rate and short consumption time.

Key words: principal component analysis ( PCA) method, Retinex enhancement algorithm, remote sensing image scene, mean shift segmentation calculation, K-singular value decomposition method (K-SVD), convolutional neural network