Abstract: The key point of face recognition is feature extraction, which was mainly described by a perfect lower dimensional feature space to image the high dimensional in the past, but recently, had been provided a new direction by the deep learning model. In this paper, a deep subspace model that under the Gabor feature descriptor modulation is put forward. The modulation based on a new type deep learning framework uses Gabor filter to process images, draws a multilayer network from the deep feature extraction built,gets the deep abstract feature modulated by the Gabor. Firstly, we compress the traditional 40 Gabor filters of 8 directions and 5 scales in scale into 8 basic Gabor filter banks; Then we put the description features filtered by Gabor into the deeply reformed subspace model separately to get the deep feature representation of the image; Secondly, deal these features with hash code, histogram block as the description features. We discuss the improvement of the face recognition by adding the deep multilayer subspace feature extraction modulation based on the FERET, ORL, CMU_PIE database. The results show that this algorithm could make a better discrimination, good robustness on the illumination, expression, and posture. Meanwhile, it also makes up for the shortcoming of the shallow layer network who easily affected by the training image.