面向无标记显微细胞图像增强的混合自适应多尺度感知驱动网络OA

Aiming at the problems of the lack of standard paired data in label-free microscopic cell images,as well as the low contrast,insufficient brightness and blurred cell edge details in labeled microscopic cell images,a hybrid adaptive multi-scale perception-driven network is proposed for label-free microscopic cell image en-hancement.First,a hybrid adaptive Retinex module is designed to perform illumination optimization and struc-tural preserving preprocessing for label-free microscopic cell images.Second,a multi-scale illumination decom-position module is emplyed to separate high-frequency and low-frequency regions of the illumination compo-nent,obtaining illumination sensitivity coefficients and local noise levels of the image.Third,a dynamic Gam-ma parameter prediction module is constructed based on a lightweight CNN to generate a spatially variant Gam-ma correction mechanism,enabling pixel-level adaptive correction.Finally,an adaptive noise detection en-hancement module dynamically selects enhancement strategies according to the entropy value of the reflection component,effectively suppressing noise amplification.Experiments results show that the proposed method outperforms existing enhancement methods in terms of key quantitative metrics(NIQE,MV,IE),verifying its effectiveness in improving the visual quality of label-free microscopic cell images.