<p>Semantic segmentation is a low-level visual task that requires denser image information and target details. However, target segmentation continue to be hindered by incomplete problems, attributable to sparse spatial features and localization information. Furthermore, the conventional non-linear activation function is prone to vanishing gradient and dying neuron problems in the neural network, which have a detrimental effect on the overall performance of the semantic segmentation task. In this paper, we propose a novel multi-feature fusion network with a detailed reinforcement module and Sarsh activation function (MFNet). The detailed reinforcement module can fuse multi-level features in the encode stage, which captures denser target details and richer visual information. To alleviate the vanish gradient and dying neuron problems, we introduce a Sarsh activation function, which is a differentiable, non-saturation, and nonlinear function with bounded below and no bounded above. It can ensure the normal gradient flow and reduce data offset. Extensive experiments are conducted to evaluate our method on three public databases, and quantitative and qualitative results indicate that our method achieves better segmentation performance than other existing methods, which also demonstrates the effectiveness and reliability.</p>