In response to the issue of excessive reliance on professional equipment for detecting air PM2.5 concentration, an air quality detection algorithm based on multi-source information and image feature generalization is proposed. Firstly, the EfficientNet-B0 was used as the backbone network for feature encoding of the input atmospheric visible image, the multi-source meteorological information, such as temperature, humidity, wind speed, pressure and light intensity, was mapped into feature vectors corresponding to the atmospheric image, and fused with the atmospheric image features. Then, the global features were output as scalars using a fully connected layer, and the PM2.5 concentration in the air was detected using a loss function. Finally, the features of atmospheric images at different scales were randomly generalized and enhanced in the training phase of the network model to enrich the sample distribution space, making the network to learn more features from limited data samples, thereby effectively improving the performance of the detection network. The experimental results show that the designed air quality detection method has higher detection accuracy and stability compared to several mainstream methods, the RMSE and R-squared obtained on the test set are 21.55 μg/m3 and 0.923, respectively. The average error obtained by detecting 8 scenarios is only 5.2%, and the maximum error is only 7.6%, which can adapt to air quality testing tasks in various extreme atmospheric pollution environments.