Few-shot semantic segmentation aims to perform semantic segmentation with a small number of samples, offering new solutions for practical applications constrained by sample limitations. The prevailing few-shot segmentation approaches typically employ whole support images as input. However, when target objects are small, their representations are easily overwhelmed by redundant background cues. Such feature interference substantially impedes the accurate segmentation of target objects in query images. To address the issue, we consider the masked support images as complementary internal information to the full support images, where the masked support images are obtained by removing irrelevant backgrounds using the support masks. Hence, we propose a simple method to fuse the support information corresponding to the two support images, which generates fused middle support features and fused query prior masks to enable more reliable prediction of query images. Furthermore, to address the issue of intra-class differences between query and support images, we propose a co-matching module that leverages query and support features to generate more representative co-support features, thereby enabling pixel-level matching between query and co-support features. Experimentally, our model outperforms previous few-shot segmentation methods on the popular few-shot segmentation benchmark dataset, i.e., Pascal- \(5^i\) and COCO- \(20^i\) , without the need to retrain the model in the 5-shot setting.