Convolution layer and attention mechanism are two of the most important foundations for the success in image denoising through neural networks. Success of attention mechanism depends on the similarity score calculated from key and query generated from features of the image. As noise level increases, effectiveness of attention score drops because similarity matching becomes difficult. To overcome this performance reduction, traditional patch-based approaches performs matching in a small neighborhood to reduce noise interference. However, neural networks calculate attention score globally. This means that effect of noise interference is more severe in attention mechanism than traditional patch-based approaches. Therefore, in this paper, we propose to add a region constraint mask in attention calculation. With this newly added region constraint, attention score is calculated from a limited neighborhood around the considered pixel instead of from the entire image, thereby, reducing interference at high noise level. In addition, a convolutional path is added to complement the new region-constraint attentive path and these two parallel paths are fused together by another neural network to generate the final denoising result. With these two strategies, experimental results show that our proposed new network architecture performs better than state-of-the-art blind denoisers both quantitatively and qualitatively.