Image-based Particle Pollution Estimation Using Convolutional Neural Network

Air quality issues have a greater impact on public health and daily behavior.Traditional air quality testing is mostly done by professional environmental testing stations,but it is limited by human resources,material and financial resources.With the widespread deployment of image acquisition equipment and the continuous innovation of machine learning algorithms,image-based particle pollution estimation using convolutional neural network can not only compensate for the geographical limitations of the inspection sites and the inability to update data in real time,but also save costs and human resources.In this paper,the following research work is carried out for the task of image-based particle pollution estimation using convolutional neural network:(1)Aiming at the problem that the existing image-based particle pollution estimation models do not achieve effective feature enhancement based on the correlation between image quality and air quality.Thus,a gradient saliency feature enhancement algorithm for image-based particle pollution estimation is proposed.Firstly,a gradient feature enhancement method based on dual-branch network is proposed to achieve the enhancement of gradient features.Secondly,the gradient saliency attention mechanism is introduced,and a feature enhancement method based on gradient saliency is proposed,so that important gradient features dominate the model training process.Through experiments with the benchmark CNN models as the image-based particle pollution estimation network,it is fully proved that the use of gradient prior information to participate in the training of the network can effectively improve the performance of the model.(2)Aiming at the problem that the existing mature models are not suitable for feature extraction in particle pollution estimation task,a shallow hybrid convolutional neural network(SHCNN)suitable for the current task is proposed.For the first time,the application of dilated convolution to the task of image-based particle pollution estimation can make up for the shortcomings of traditional convolution in processing image information with small receptive fields and environment-dependent feature extraction.At the same time,a lightweight global channel attention module(GCA)is proposed to improve the model’s attention to effective features during network training.During the model training,a new objective loss function named FMSE is proposed to help better complete the image-based particle pollution estimation task.(3)Aiming at the limitation of using only images to evaluate air quality,an image-based particle pollution estimation algorithm that integrates meteorological information is proposed.Firstly,a method for encoding meteorological information based on word embedding is proposed,which can obtain effective low-dimensional dense vector expression form of meteorological information in the model training process.Secondly,a multi-feature fusion network is constructed to fuse meteorological coding features and image features.Experiments show that using the method in this paper to integrate the four types of meteorological features of temperature,humidity,pressure,and wind speed with image features can effectively improve the robustness and versatility of the model.The research work in this paper shows that using gradient information to achieve feature enhancement,the construction of a shallow network for feature extraction,and the integration of meteorological information to increase feature diversity can all improve the performance of the image-based particle pollution estimation model.