Image-based Detection Of Air Pollution Levels Of Fine Particulate Matter

Air pollution brings serious harm to human health and limits people’s life and travel.Effective air pollution monitoring is great importance to human health and travel.The current main way to obtain air pollution information is to rely on air quality monitoring stations to determine pollutant concentrations,which are then converted into an air pollution index for public inquiry.In addition to the disadvantages of high instrumentation cost,coarse-grained detection mode and poor real-time performance,this method also fails to obtain accurate air pollution information in areas where air quality monitoring stations are not installed.The images,on the other hand,contain intuitive information,and the use of images to identify the level of air pollution can achieve low-cost,fine-grained and real-time convenient detection.In this article,an image-based method for detecting air fine particle pollution levels is investigated by collecting image data.This work combines computer vision and deep learning algorithm.The main work is as follows.(1)A two-way residual network structure-based air fine particulate matter pollution level detection model was constructed.The primary network performs feature extraction on the original image,and the secondary network is trained on the dark channel information to enhance the data.A pre-parallel fusion strategy is used without increasing the dimensionality of the features,which reduces the parameters and computation while enhancing the feature information.Compared with using a single residual network structure,its detection accuracy is improved by 8.8%.(2)Adding attention mechanisms to the main network of the constructed detection model.Adding channel and spatial attention mechanisms to the network structure to strengthen or suppress feature information and focus global key information.In the selection of the loss function of the network model,the cross-entropy loss function is improved by adding a dynamic adjustment factor to reduce the weights of easily classified samples,so that the model can focus more on the hard-to-classify samples during training and thus tap into the feature information that is difficult to extract.The experimental results show that the accuracy of this model can reach 76.2%,which is 7.7% and 2.2% better than ResNet+VGG(ResidualNetworks+Visual Geometry Group)model and ResNet+SCA(ResidualNetworks+Spatial Channel Attention)model,respectively.The improvement is 7.7%and 2.7%,respectively.(3)An air fine particle pollution level recognition model incorporating multiple features is constructed.The RGB images were converted to the HSI channel,which is more compatible with human visual characteristics.The differences of each pollution level image in H,S and I channels are analyzed,and the image processing is performed by using pixel-level feature fusion to obtain a new feature map.The experimental results show that this method can identify the air fine particle pollution levels more accurately.Its highest accuracy can reach 77.9%,which is 9.6% and 4.6%better than ResNet+VGG model and ResNet+SCA model,respectively.The effectiveness of the model is verified after testing in real scenarios.