Mixed Pruning Method for Vehicle Detection

摘要

Object detection is a popular direction in computer vision and digital image processing, and has important implementation significance in many fields. In unmanned driving, it is also an indispensable component, and vehicle detection as its important branch has also received much attention. With the continuous development of deep learning in recent years, convolutional neural networks have been widely used for object detection. Compared with traditional detection methods, it has better generalization ability and recognition accuracy. However, most current deep learning-based object detection methods rely on the graphics processor (GPU) to ensure the real-time detection. This is difficult for an unmanned system with limited memory and computing power resources to support its operation. In order to solve the problem of excessive resource consumption of these hardware platforms, we propose a vehicle detection algorithm based on a hybrid pruning model, which ensures that the accuracy of the model is within a controllable range while reducing the number of model parameters as much as possible. The model increases the sparsity of the network by using the L1 norm regularization method, and cyclically prunes the model through channel pruning and layer pruning. The experimental results show that the compressed network model is reduced by 83% compared to the original model, and the corresponding speed is reduced to half of the original.