A spectral-spatial feature extraction method with polydirectional CNN for multispectral image compression

Convolutional neural networks (CNN) has been widely used in the research of multispectral image compression, but they still face the challenge of extracting spectral feature effectively while preserving spatial feature with integrity. In this article, a novel spectral-spatial feature extraction method is proposed with polydirectional CNN (SSPC) for multispectral image compression. First, the feature extraction network is divided into three parallel modules. The spectral module is employed to obtain spectral features along the spectral direction independently, and simultaneously, with two spatial modules extracting spatial features along two different spatial directions. Then all the features are fused together, followed by downsampling to reduce the size of the feature maps. To control the tradeoff between the rate loss and the distortion, the rate-distortion optimizer is added to the network. In addition, quantization and entropy encoding are applied in turn, converting the data into bit stream. The decoder is structurally symmetric to the encoder, which is convenient for structuring the framework to recover the image. For comparison, SSPC is tested along with JPEG2000 and three-dimensional (3-D) SPIHT on the multispectral datasets of Landsat-8 and WorldView-3 satellites. Besides, to further validate the effectiveness of polydirectional CNN, SSPC is also compared with a normal CNN-based algorithm. The experimental results show that SSPC outperforms other methods at the same bit rates, which demonstrates the validity of the spectral-spatial feature extraction method with polydirectional CNN.