The existing sparse projection view-based cone-beam computed tomography(CBCT) reconstruction methods suffer from low image quality, prolonged reconstruction times, and an inability to perform end-to-end reconstruction. To address these issues, this paper proposed a deep filtering multi-projection angle feature extraction network(FMA-Net). Initially, it processed projection data in the frequency domain to effectively suppress noise and artifacts. Subsequently, it introduced a multi-projection attention module and a visual state space module to enhance the networks capability to extract features from projection data, and improving the utilization rate of projection feature information. Finally, it employed a multi-projection angle feature interaction module to obtain similar information from the same projection point at different projection angles, thus enhancing the quality of reconstructed data points in CBCT. Comparative experiments were conducted on 21 real walnut datasets and 1 018 pulmonary CT projection datasets with FMA-Net and six conventional methods, such as FDK, SART, SART_TV, CGLS, CNCL and DIF-Net. The results show that FMA-Net outperforms these methods under four different projection image quantity conditions of 18, 21, 24 and 27. Specifically, on the walnut dataset, FMA-Net achieves an average RMSE reduction of 14.6%, an average PSNR increase of 4.3%, and an average SSIM increase of 1.75%. On the LIDC-IDRI pulmonary dataset, it shows an average RMSE reduction of 16.3%, an average PSNR increase of 5.4%, and an average SSIM increase of 5.5%, while also leading in reconstruction speed. All results indicate that FMA-Net can rapidly reconstruct high-quality cone-beam CT images from sparse projection views.