In recent years, deep learning-based semantic segmentation has seen extensive application in areas like autonomous driving and remote sensing image processing. Numerous models aim to boost segmentation accuracy and efficient information extraction, typically using a convolution network structure for better model generalization. But they seldom pay attention to the scalability of image information and noise generation during information transmission. However, it is often difficult to optimize segmentation details. In our approach, we modify the convolution structure to fuse features from various scales and integrate additional image features with optimized information transfer to enhance accuracy. At the same time, our method does not rely on a unique network structure, which improves generalization. We analyze feature detail transfer in the convolution structure and propose two strategies to improve feature processing. Firstly, we identify causes for the loss of image detail and noise spatial distribution. Subsequently, we develop two universal methods to improve information transmission. One uses image gradient information to guide boundary segmentation and address segmentation challenges at complex boundaries. The second filters convolution results to suppress noise transmission through the network.