Abstract: A sparse reconstruction method for Head-Related Transfer Functions (HRTFs) based on directional feature extraction is proposed. This method addresses the limitation of existing deep learning-based HRTF sparse reconstruction approaches, which rely on HRTF inputs sampled from specific grids. It takes spatially sparse HRTFs and their directional information as input, extracts directional features through a directional encoder, and separately integrates them with the frequency domain features of the spectral encoder and the temporal features of the onset time encoder. This fusion enhances the model’s understanding of the relationship between sparsely sampled HRTFs from various directions and high-resolution HRTFs, thereby improving its adaptability to variations in input directions. The experimental results indicate that the proposed method outperforms existing methods in reconstructing full-space HRTF under non-specific grid sampling conditions. It also demonstrates superior performance in objective evaluation metrics such as spectral distortion and interaural time differences, as well as in the accuracy of auditory localization models and virtual sound source localization tests.