As the primary pollutant in China's urban atmosphere, PM

$_{2.5}$

poses a great threat to the health of residents and ecological stability. Efficient and effective PM

$_{2.5}$

concentration monitoring is essential. Nonetheless, the popular devices for PM

$_{2.5}$

monitoring are developed based on two standards: the micro-oscillation balance method and the

$\beta$

-ray method, which have high purchase and maintenance costs and slow calculation rates. To this end, we put forward a real-time and reliable vision-based estimation algorithm of PM

$_{2.5}$

concentration. To be specific, the proposed method first develops two natural scene statistical analysis-based visual priors to measure saturation and structural information losses caused by the ‘haze’ formed by PM

$_{2.5}$

. Moreover, we develop a lightweight deep belief network (DBN)-deep neural network (DNN)-based PM

$_{2.5}$

concentration estimation model, which learns the mapping from the designed visual priors to PM

$_{2.5}$

concentrations. Experiments confirm the superiority of our vision-based PM

$_{2.5}$

concentration estimation method by comparison with state-of-the-art photo-based PM

$_{2.5}$

monitoring methods.