Abstract: In order to improve the vehicle detection result under multi-weather conditions， a convolutional network based on modified knowledge distillation method was proposed. The network uses cumbersome CNN（Convolutional neural network） as teacher network and lightweight CNN as student network. Without adding new training dataset and slightly increasing the number of light network parameters， the performance of the lightweight CNN under multi-weather vehicle detecting conditions can be improved. The network utilizes a specialized data enhancing method to generate a multi-weather feature dataset. The teacher network is trained on the data without weather feature， and the student network data is trained simultaneously on the data with weather features. Considering that images without weather features can provide more information relatively， through this training method， the student network can better learn the information generated by the teacher network. Finally， through the multi-weather vehicle detecting performance of the network trained and tested on BDD100k dataset with enhanced weather dataset， the detectability and stability of the student model in multi-weather environment boosts. The comparison test of the generalization ability of multiple networks is carried out on DAWN multi-weather dataset， and the modified distillation convolutional network achieves certain advantages in average precision and detection speed.

Key words: machine vision, deep learning, knowledge distillation, convolutional neural network, object detection, data enhancement