Improved Weighted Map and Object Loss Function Based on Saliency Detection Algorithm

Salient object detection (SOD) pre-processes downstream computer vision tasks. Based on the Saliency and Detail Map Interactive Model for Salient Region Detection, we constructed detail maps. In this study, the loss function for training models was refined. The original loss function is not tailored for detail maps with inadequacies. Thus, we incorporated a non-linear function into the weighted maps of the weighted Binary Cross Entropy (weighted BCE) and redirected weighted maps in edge regions for SOD prediction. By opting for the S-measure loss over the IoU loss, model performance was enhanced. Compared with the original method and other eleven state-of-the-art (SOTA) methods, the proposed approach outperformed other methods for three evaluation metrics.