An Online-Updating Deep CNN Method Based on Kalman Filter for Illumination-Drifting Road Damage Classification

Abstract

Damage of road surface, e.g., Cracks, is the critical problems in road maintenance. Previous automotive road damage detection methods mainly focus on hand-crafted features and shallow classifier models. Recently, deep learning methods have also been proposed. The deep neural networks consist of dozens of parameters, which is usually optimized by the Mini-batch Stochastic Gradient Descent Algorithm (MB-SGD). However, MB-SGD is awkward for online update when new training samples from a drifting system condition, e.g., illumination, are received. In this paper, we first present an experimental study on how the illumination change affects the generalization of a pre-trained deep convolutional neural networks. Then, we propose a novel Kalman Filter based method for online updating the network parameters. Experimental results convince that the illumination change can affect the performance of a pre-trained CNN using training samples from a fixed illumination condition. By using the proposed method, the CNN can online adapt its parameters in the classifier layer to the received training samples sequentially, which leads to a better classification performance. The proposed method alleviates the need of huge amount of training samples covering all system conditions, which are hard to collect and costly.