Compressing Deep Neural Network for Facial Landmarks Detection

State-of-the-art deep neural networks (DNNs) have greatly improved the performance of facial landmarks detection. However, DNN models usually have a large number of parameters, which leads to high computational complexity and memory cost. To address this problem, we propose a method to compress large deep neural networks, which includes three steps. (1) Importance-based neuron pruning: compared with traditional connection pruning, we introduce weights correlations to prune unimportant neurons, which can reduce index storage and inference computation costs. (2) Product quantization: further use of product quantization helps to enforce weights sharing, which stores fewer cluster indexes and codebooks than scalar quantization. (3) Network retraining: to reduce training difficulty and performance degradation, we iteratively retrain the network, compressing one layer at a time. Experiments of compressing a VGG-like model for facial landmarks detection demonstrate that the proposed method achieves 26x compression of the model with 1.5% performance degradation.