TSDSNet: a three-branch detail-sensitive and noise-suppressed fusion network for label-free microscopic image enhancement

Label-free microscopic cell images often suffer from low contrast and inadequate brightness, leading to detail loss and edge blurring. This study proposes a zero-reference learning-based triple-branch fusion network for label-free microscopic image enhancement. The network incorporates a detail-sensitive enhancement module to extract multi-scale gradient features, improving edge and fine-detail preservation. A multi-scale Haar wavelet downsampling module is embedded to efficiently separate illumination, reflection, and noise components in the frequency domain. Gamma correction enhances brightness and contrast in the illumination component, while a noise-suppression differential operation refines the reflection component before fusion via Retinex theory. Additionally, a multi-constraint loss function combining Retinex reconstruction loss, detail-sensitive loss, and illumination-guided noise suppression loss is proposed to enhance structural details while suppressing noise amplification. Experiments on multiple public datasets demonstrate that the proposed method outperforms existing state-of-the-art techniques across various image quality assessment metrics, validating its efficacy for microscopic image enhancement.