A classification method of fog image for USV visual system

Abstract

Video image defogging algorithms cannot intelligently identify whether current scene has fog or not. In order to improve the intelligent defogging ability of unmanned surface vessel (USV), a classification method based on multi-feature extraction is proposed to distinguish whether a surface image is a fog image or not. A surface image sample library is first established and 8 features that are significantly different in fog and clear images are extracted. Second, the features are normalized separately, and the Back Propagation (BP) neural network and Support Vector Machine (SVM) classification method with different combinations of features are used to classify the surface fog and clear image. Experimental results show that BP neural network is better than SVM for surface fog image classification with limited samples. And, when using BP neural network for classification, we only need 3 features of improved mean, visibility and image intensity. With these features, the average recognition rate can reach 98.75%. Furthermore an experiment shows that the extracted features used in our paper can also be as classification features of fog image concentration.