A better way to monitor haze through image based upon the adjusted LeNet-5 CNN model

Abstract

Recognition of haze images is a prerequisite for appropriately realizing image dehazing algorithms, which play an important role in detecting various types of outdoor environments. In this study, the recognition of haze images is performed by adjusting the parameters and structure of the classic LeNet-5 model. The basic adjustments include the number of feature extractions and the convolution kernel size, which can increase the number of receptive fields and enhance the detailing and contrast. Simultaneously, different optimizers are used in our method for comparison; the Adam optimizer was found to be the best. Through rigorous experiments, the ameliorated model is shown to improve the recognition accuracy to 98.308%. At the same time, with the shallow learning method, fewer parameters, and lesser equipment requirements, the applicability of the method is improved. The image recognition technology is applied to a haze image field, which shows good performance. The realization of haze image recognition not only increases diversification of applications, but also provides a new and efficient detection method for urban disaster warning, which is of great significance.