【摘要】 针对结直肠图像中息肉尺寸大小不一、边界模糊以及内窥镜图像清晰度受限等问题，提出了一种基于内容引导异构双解码器的特征融合网络(HCGFNet)。HCGFNet中编码器网络采用异构多路径自适应特征融合模块(HAF),通过异构多数据流更精准地捕获复杂肠道环境中隐匿的各类小型息肉与周边特征信息。解码器网络采用内容引导特征融合注意力机制(CGFA),逐层处理解码阶段特征图中干扰信息，细化目标边缘分割效果并辅助重建灰度图。最终分别在KvasirSEG、CVC-ClinicDB、CVC-ColonDB息肉数据集上进行广泛对比，结果表明，所设计的HCGFNet相较于目前主流模型，在各项性能中均有提升。引入HAF、CGFA模块后，各项性能较基准模型提升2%～5%,较最先进模型提升1%～2%。更多还原

【Abstract】 In response to issues such as varying polyp sizes, blurred boundaries in colorectal images, and limited clarity in endoscopic images, this paper proposes a feature fusion network based on content-guided heterogeneous dual decoders(HCGFNet). In HCGFNet, the encoder network employs the proposed Heterogeneous Multi-path Adaptive Feature Fusion module(HAF), which captures concealed small polyps and surrounding feature information in complex intestinal environments more accurately through heterogeneous multi-data streams. The decoder network utilizes the designed Content-Guided Feature Fusion Attention mechanism(CGFA) to process interfering information in feature maps during the decoding phase layer by layer, refining the segmentation of target edges and assisting in grayscale image reconstruction. Extensive comparative experiments conducted on the Kvasir-SEG, CVC-ClinicDB, and CVC-ColonDB polyp datasets demonstrate that the proposed HCGFNet outperforms current mainstream models across various performance metrics. After incorporating the HAF and CGFA modules, performance improvements range from 2% to 5% compared to the baseline model and from 1% to 2% compared to state-of-the-art models.更多还原