The detection and recognition of long-distance moving targets on space-based platforms face challenges such as wide bandwidth， blurry images， and small target sizes. To address these issues， a multi-scale， multi-stage convolutional neural network incorporating an attention mechanism is proposed， meeting the demands for high real-time performance， generalization， and deblurring quality. The approach employs a lightweight multi-scale， multi-stage network with an optimized module count to ensure real-time processing， integrating a frequency-domain-based self-attention solver， a discriminative FFN （F-D）， and a convolutional block attention module （CBAM） to extract critical spatial and frequency domain information. The experimental results show a deblurring restoration rate exceeding 34 frame/s， with time consumption reduced to one-third of conventional methods. On the infrared small target dataset， the PSNR exceeds 32 dB and the SSIM surpasses 0.87， while on the visible light small target dataset， the PSNR exceeds 17 dB and the SSIM surpasses 0.93.The algorithm demonstrates strong generalization across wideband scenarios， effectively restoring the contours and shapes of small targets.