YOLOSA: Object detection based on 2D local feature superimposed self-attention