Abstract: Shadow makes many computer vision tasks perform poorly, and the complex relationship between the shadow and scene makes it a research hotspot. For the shadow area obtained by learning the features of shadow with the help of artificial mask and applying it to shadow removal, the cost of manual annotation is higher. By learning the method of direct mapping from a shaded image to an unshaded image, there may be problems of producing wrong mapping and interfering with non-shaded regions that should not be involved in the mapping. Therefore, a separation and reconstruction network is proposed, which uses the idea of independent component analysis to separate the shadow features from the environment features, and carries on the difference mapping to the separated features, so as to achieve shadow removal. A large number of experiments were carried out using AISTD and SRD datasets, and indicators were compared by RMSE. The results reveal that the proposed algorithm, implemented without the use of masks (SRNet), demonstrates a noteworthy 17.5% reduction in average error within shadowed areas compared to mask-free methods, resulting in a 2% decrease in overall areas. In contrast, the approach employing masks (mSRNet) exhibits commendable comparability with its masked counterparts. The algorithm exhibits superior efficacy in the context of shadow removal.