Lightweight smoke identification method based on spatial-temporal feature fusion

The invention discloses a lightweight smoke identification method based on spatial-temporal feature fusion, and the method comprises the steps: firstly constructing a lightweight deep learning network (TSNet) of spatial-temporal feature fusion, and the network comprises a shallow spatial feature extraction module and a spatial-temporal feature extraction module; and on the basis of the network, a lightweight smoke identification model based on spatial-temporal feature fusion is constructed. In the model, firstly, a shallow spatial feature extraction module in the TSNet is utilized to extract local spatial features of a single-frame picture, the local spatial features are abstracted, and the resolution of a feature map is reduced step by step; the method comprises the following steps: respectively extracting time sequence features among multiple frames of smoke pictures and deep abstract features of a single frame of smoke picture by using improved probsparse self-attention through a time-space feature extract