Infrared targets have no texture featurein the case of remote imaging,and are susceptible to noise and clutter.It is difficult to detect such weak feature targets.An attention based full convolutional neural network is proposed,and the multiscale target detection is regarded as a pixel wise image segmentation problem to adapt small target extraction in complex backgrounds.The time domain saliency map is constructed,and the images generated by different cameras can be projected into the same representational space.It can improve the generalization ability of the network model.The attentional mechanism is used to learn bimodal and multi level information,and to adaptively enhance the learning ability around targets.The performance comparison under different conditions is implemented based on real data and simulation data.The experimental results show that the proposed method can effectively improve detection performance for multiscale and weak feaure targets.Meanwhile,the neural network has quicker convergent rate.