Saliency Object Detection Method Fusing Global and Local Features Based on Transformer
GE Yichen
ZHANG Ming
Abstract：To address the inefficiency of existing saliency object detection methods in global and local feature fusion,a saliency object detection method,namely,vision transformer-deep transformer decoder(ViT-DTD)was proposed.Firstly,the method chunked the input image by means of a visual transformer which first captured local features at a shallow level and then aggregated global features at a deep level.Secondly,feature information was ensured to be fully utilized in the decoding process by fusing shallow local features and deep global features layer by layer,combined with a dense connectivity strategy.Finally,the fused features were decoded by layer-by-layer upsampling to recover the high-resolution saliency map of the image.The results showed that the mean absolute error(MAE)of the method on the two benchmark test datasets was reduced by an average of 14.0％compared with the label decoupling framework(LDF)method,the weighted F metric was improved by an average of 3.7％,and the structure-measure(S)index was improved by an average of 2.7％.Compared with the other methods,the ViT-DTD has better performance in complex scenarios,and it provides a more accurate and stable saliency detection scheme.
Keywords：saliency object detectiontransformerlocal featureglobal featureViT-DTD
Publication Date：2024-12-20
Online Publishing Date：2025-08-15(First online date of this platform, not the publication date of the document)
Pages：6( 464-469 )