A Traffic Target Detection Algorithm Based on Hybrid Model Compression Method

Abstract:
With the development of the computer vision field, convolutional neural networks are widely used in object detection algorithms. However, the deep learning network model has high computational complexity, occupies a large storage space, and is difficult to deploy on low-computing devices. Aiming at the difficulty of deploying traffic target detection algo-rithms in practical scenarios, this paper proposes a target detection network compression method based on fusion of pruning and quantization. First, based on the Yolo series target detection algorithm, we perform an unstructured pruning operation with a variable threshold for the network model to be optimized, which can remove more information than the structure and reduce the parameters of the network model. Then, the network is quantized into 8-bit integer data through quantization-aware training to compress the network size. Experiments show that the optimized network model has less accuracy loss after variable threshold pruning and data quantization, and the model size is greatly compressed, which can achieve high detection capability even in the case of limited resources.