In clinical medicine, colonoscopy is an important screening tool for colorectal cancer. Accurate segmentation of polyps during colonoscopy plays a critical role in the early detection and treatment of cancer. However, the diversity of polyp images in different regions of the colon increases the difficulty of polyp segmentation, making it a challenging task. To address this challenge, we propose a new polyp segmentation network based on an attention mechanism. Specifically, we place great emphasis on capturing the diverse features of polyps and exploit the capabilities of the middle layer of the network for feature extraction, thereby enhancing our ability to effectively characterize polyps. To achieve this, we have introduced a Feature Extraction Module (FEM) dedicated to enriching the image representation, resulting in improved segmentation accuracy. Extensive experiments demonstrate that our segmentation network performs excellently on five challenging datasets.