Discrete cosine single-pixel imaging (DCSPI) is a high-efficiency single-pixel imaging (SPI) technique that is endowed with energy compaction and computational simplicity by the discrete cosine transform (DCT). The image can be efficiently retrieved using the inverse discrete cosine transform (IDCT). However, when the sampling ratio is low (below 6.25 %), numerous DCT coefficients are absent, which poses challenges in obtaining high-quality images. To estimate the absent DCT coefficients, we incorporate the frequency information features into a deep convolutional network architecture and use a frequency loss function in the DCT domain for network training. Thus, the imaging quality of DCSPI under low sampling ratio could be improved. The simulation and experimental results validate the effectiveness of our proposed method.