Fine-grained Image-Text Feature Alignment for large Vision-Language Models