WKCN:基于小波-KAN协同优化的自适应图像描述算法OA

This study addressed key challenges in image captioning,including inadequate multi-scale texture representation,redundant feature fusion,and limited dynamic semantic modeling.This paper proposed a novel algorithm named WKCN based on wavelet-KAN collaborative optimization.It designed the wavelet-KAN multi-scale nonlinear enhancement(WKMNE)modu-le to decompose image features using Daubechies-4 wavelet bases and enhance textures via B-spline interpolation in KAN.A KAN-based adaptive feature fusion(KAN-AFF)mechanism dynamically generated spatial and channel weights to integrate global features from ResNet50 and frequency-domain features from WKMNE.Finally,a KAN-enhanced dynamic decoder(KED)replaced the static feed-forward network in Transformer with a learnable KAN activation module to strengthen semantic mapping.Experiments on the MSCOCO dataset show that WKCN achieves optimal scores on BLEU-1(81.1),ROUGE-L(59.1)and CIDEr(133.6).Ablation studies confirme the synergy between multi-scale feature extraction and dynamic deco-ding.Hyperparameter analysis verifies the rationality of parameter selection.Cross-dataset tests on Flickr30k and NoCaps demonstrate strong generalization capability.Visualization analyses illustrate the model's effectiveness intuitively.This work provides a verifiable nonlinear optimization paradigm for cross-modal semantic generation tasks.