Scene text detection, which localizes text in images, is key to scene text reading. While convolutional neural networks have advanced this field, challenges persist due to diverse scales, irregular shapes, wavy and rotated text, scale variations, different fonts, noisy images, complex backgrounds, and extreme text aspect ratios. The previous approach, despite its advantages, struggles with long-range dependency modeling. This is crucial for understanding connections between distant elements in the data. For example, in scene text detection, text often spans a large image area, requiring efficient long-range modeling, a challenge with big datasets or complex structures. In this paper, we propose TextCSCN (Text Contrast Self Calibrated Network). Our innovation stems from the integration of supervised contrastive learning into the scene text detection method. We introduce self-calibrated convolution to address long-range dependencies. Specifically, our method uses SCNet-50 as a backbone and achieves an F-measure of 88.43 on the Total-Text dataset. Notably, we have selected and incorporated a range of technologies that are particularly suitable for improving scene text detection.