基于Decoupled-FR Net的伪造图像检测模型OA

To address the challenges of fine-grained feature perception and cross-region dependency modeling in forged image detection,a detection model named decoupled frequency refinement network(Decoupled-FR Net)was proposed.The model was designed with a spatial-channel decoupled attention mechanism,which effectively avoided the feature coupling problem in traditional hybrid attention and enhanced the independence and discriminative power of feature representation.A feature refinement module was introduced to enhance the perception ability of subtle tampering through hierarchical feature calibration and fusion.Combining with a context-aware mechanism,long-distance dependencies across regions were captured,thereby improving the overall detection performance.The results showed that,the accuracy and average precision of the Decoupled-FR Net model on the forensic synthetics(ForenSynths)dataset were improved by 2.4 and 0.5 percentage points,respectively,compared with the inter-patch dependency network(IPD-Net)model,and on the generative adversarial network(GAN)generation detection(GANGen-Detection)dataset,average precision was improved by 0.1 percentage points,compared with the frequency domain network(FreqNet)model.The model provided a new solution for fine-grained forged image detection and was of important application value in the field of multimedia forensics.