A Lightweight Air Quality Monitoring Method Based on Multiscale Dilated Convolutional Neural Network

Air quality monitor is important for both the environment and human health. In recent years, many computer vision-based air quality monitoring methods have been proposed and have achieved good accuracy in specific scenarios. However, current deep-learning-based models' deployment on mobile devices is impeded due to their excessively large size. To build a lightweight model, we investigate the potential of dilated convolutional kernels, which can effectively increase the receptive field and acquire multiscale information while maintaining model size. Eventually, we propose a novel method termed multiscale dilated network. The network is tailored to the characteristics of the air quality dataset by rearranging the dilate rate of the dilated convolutional kernels to focus on the global and semantically higher level features, thus improving the detection efficiency and accuracy of the model. Our proposed multiscale dilated block also includes a multiscale fusion mechanism that enables the model to extract air quality information at multiple scales simultaneously. Compared to the current state-of-the-art models, the multiscale dilated network integrates three different scales of dilated kernels, which leads to a reduction of 86.7% of parameters and 88.5% of floating point operations and achieves a top-1 classification accuracy of 94.2%. With the proposed model enhancements, we envision a future where mobile phones and IoT edge devices can function as mobile air quality monitoring stations, thereby creating a dynamic and distributed air quality map for air quality protection.