Object-based audio content is becoming popular to represent immersive spatial sound because audio object positions can be totally independent of the locations of available loudspeakers. The encoder of the spatial audio object coding (SAOC) method transmits a mono downmix signal with spatial parameters of all audio objects, and the decoder reconstructs the original audio objects. However, the decoded audio object by SAOC has severe aliasing distortion, especially at low bitrates. This study focuses on distortion reduction of decoded audio objects at low bitrates. We present a new audio coding method via nonuniform fine sub-bands division strategy with convolutional autoencoder and a dense convolutional network mixture model, aiming to effectively compress spatial parameters of all audio objects. Evaluations indicate that the proposed approach reduces the aliasing distortion by 2.02%. It has achieved the best performances in both objective and subjective results with a low bitrate 1 kb/s per object.