Deep learning applied to seismic First Break (FB) picking has been developed for many years, and numerous researchers have used Image Semantic Segmentation Networks (ISSNs) for multi-channel FB picking. The existing seismic FB picking method based on ISSNs usually adopts two label calibration and FB determination methods, one is to divide the seismic signal into pre-FB and post-FB, and pick up the FB through mask segmentation; the other is to divide the seismic signal into FB and non-FB, and pick up the FB by extracting the highest confidence point of each trace. The former suffers from FB false pickup with localized regional continuity due to the mask edge blurring problem; The latter, because of the large proportion of positive and negative samples, tends to make the network hard fitting and cannot be applied to data with complex FB waveforms and large size. Based on this, a dual-channel mask interaction seismic FBs picking method is proposed, which ensures the network's FBs feature recognition ability by banded FBs range mask, and enhances the network's FBs accurate picking ability by linear preferred FBs mask, which effectively avoids the shortcomings of the existing methods. Theoretical experiments show that the method has good noise resistance and can be generalized to higher noise level data by training in low noise level data. When the method is applied to the field data, it achieves higher FB picking accuracy than the existing methods, and the number of traces with picking error of 0 ms is as high as 75.9%, which is 7.1%, 26.8%, and 15.6% higher than that of STUNet, SegNet, and Res-Unet, respectively, and greatly improves the efficiency of high-quality seismic FB picking. Meanwhile, the approach adopts a lightweight network model with high inference efficiency and easy engineering deployment, which has practical application value.