Semi-supervised semantic segmentation (SSS) aims to alleviate the burden of time-consuming pixel-level manual labeling. Current mainstream methods typically adopt a teacher-student framework based on pseudo-labeling, where the teacher model generates pseudo-labels to guide the training of the student model on the unlabeled data. However, these pseudo-labels commonly contain noise, which leads to accumulated loss during training and significantly degrades the model performance finally. Furthermore, simply discarding low-quality predictions is improper for resulting in inefficient training samples utilization. To address the above issues in SSS, we propose an Enhanced Teacher-Student Framework with Weighted Loss and Presence Awareness (TeSuLoPA), which improves the quality and utilization of pseudo-labels through collaborative optimization. Initially, a Local Enhancement Fusion Encoder combines enhanced local features with original features to capture comprehensive contextual information to obtain reliable pseudo-labels. Secondly, we design a Dynamic Loss Weighting strategy that dynamically assigns sample- and batch-wise weights to predictions based on their quality, enabling effective utilization of both high- and low-quality pseudo-labels. Additionally, an Auxiliary Presence Awareness model evaluates the reliability of pseudo-labels to support batch-wise loss weighting and generates foreground predictions to supply auxiliary signals for the student model. Experiments on SSS benchmarks demonstrate that the proposed TeSuLoPA outperforms most existing methods. Specifically, on Pascal VOC 2012 datasets with only 183 labels, it achieves a 2.29% mIoU gain over the state-of-the-art method.