Research on Improved Algorithm of Significance Object Detection Based on ATSA Model

Saliency detection refers to accurately positioning and extracting significant objects or regions in the image. Most effective object detection methods are based on RGB-D and adopt the dual-flow architecture with RGB and depth symmetry. At the same time, the asymmetric dual-flow architecture can also effectively extract rich global context information. The existing ATSA model uses asymmetric dual-flow architecture to locate significant objects accurately. However, this model's initial learning rate needs to be improved, and choosing a suitable learning rate takes work. Therefore, in order to improve the overall performance of the model, the RangerQH optimizer algorithm was introduced to enable the model to adjust the learning rate during the training process dynamically, and the cross-entropy loss function of the original model was replaced with a mixed loss function composed of Focal loss and Dice loss. The results show that E-measure, S-measure, F-measure, and MAE improve the seven existing public RGB-D datasets.