Air pollution measurement based on hybrid convolutional neural network with spatial-and-channel attention mechanism

Air quality is tightly correlated with human health, and long-term exposure to air pollution can pose a serious health risk to humans. In recent years, image-based air quality detection methods have been proposed and have achieved good accuracy in specific scenarios. However, most of the methods are still based on pure CNNs with fast inference speed but limited accuracy. Some also invoke a single channel or spatial attention mechanism, with improved accuracy but much slower inference speed. To have both advantages we propose the Spatial and Channel Calibration Network (SCCNet). The network combines spatial and channel attention to improve the detection efficiency and accuracy of the model by better extracting global information to focus computational resources on regions that are more important to the task. Our proposed channel averaging pooling (CAP) module significantly reduces the number of parameters in the model while extracting global information, improving the detection speed of the model. We also introduce a discrete cosine transform (DCT) method to transform images from the spatial domain to the frequency domain, which enhances the extraction of fine-grained features and improves the model’s classification ability for air quality detection tasks. Our experimental results show that SCCNet achieves an accuracy of 92.17% with about 30 million parameters in an air quality detection task, which is 1.65% and 1.71% more accurate than Swin Transformer (based on spatial attention) and SENet (based on channel attention) for a similar number of parameters. Our code and models will are publicly available at https://github.com/Fucheng-Wu/SCCNet.