Bridging the Modality Gap by Similarity Standardization with Pseudo-Positive Samples