Salient object detection (SOD) aims to focus the most prominent objects in images for many vision applications such as object tracking, image and video segmentation and also some medical image analysis tasks like tumor highlighting. Recent methods have some limitations in which most of existing approaches are designed for handling images of fixed sizes and secondly, they often fail to capture objects in complicated backgrounds. In this article, we present an Adaptive Multi-Resolution Fusion (AMRF) approach for overcoming these limitations. The approach, called as AMRF, consists of several deep learning networks, which exploit effective features at different resolutions. The AMRF framework extracts edge-preserving features using convolutional neural networks (CNNs) and encodes contextual dependencies through gated recurrent neural networks (RNNs), followed by image decomposition into full, half, and quarter resolutions. A softmax-driven energy model learns per-image scale weights γn, with an attention-guided fusion mechanism producing the fused feature map. Saliency maps are generated via a U-Net architecture optimized by a combined binary cross-entropy (BCE) and Intersection over Union (IoU) loss, while adversarial refinement using generative adversarial networks (GANs) sharpens boundary precision. An adaptive regulator dynamically adjusts the weighting parameters (α, β) to balance local and global evidence under the composite loss L total . Experimental evaluation on MSRA-B, DUTS, and HKU-IS datasets demonstrates that AMRF achieves 95.1% precision, 92.8% recall, 93.9% F-measure, and a mean absolute error (MAE) of 0.056.
Uddin et al. (Thu,) studied this question.