Real Time Dust Detection with Image Pyramid and Improved ResNeSt

Image based dust detection is an essential task for environment protection or security management in places such as sea port, flour factory or grain warehouse. In this paper, a real time dust detection framework using image pyramid and improved ResNest is proposed to rapidly calculate the distribution of dust in an image from the surveillance video data. Compared with the image classification methods such as ResNet or segmentation method such as Unet, the proposed method makes a balance between speed and precision with the help of image pyramid. Also the ResNest backbone is enhanced according to the dust image feature to achieve a better performance. The experiment results indicate that the proposed method can achieve a speed up over 3.5 times while preserve the 73% accuracy compared with mainstream segmentation methods such as UNet.