Small target detection method based on scale adaptive saliency filtering

The invention discloses a small target detection method based on scale adaptive saliency filtering. The method comprises the following steps: performing feature extraction on a training image by using a feature extraction network; the feature maps are input into encoders to be processed, each encoder layer predicts a foreground confidence map and a token filtering threshold value, and feature interaction is carried out through variable self-attention; the encoder comprises six layers, each layer calculates variable self-attention, and variable cross attention is used in the decoder to decode features to obtain a detection result; a scale coefficient rho is introduced when the saliency score is calculated, and rho is calculated according to the size of a detection frame. And calculating loss Lf according to scale adaptive significance supervision, and adding the loss Lf into the total loss. According to the method, more latitude is given to the small target through one scale coefficient, and the information ut