Correction Of Uneven Illumination In Microscopic Image Based On Flatfield Prior Approximation And Deep Learning

A desirable criterion,which microscopic images intended for biomedical researchers or computer processing and analysis should fulfill is global illumination intensity uniformity.However,in the image collected by a microscope,uneven illumination distribution is almost inevitable due to equipment or man-made reasons.Effectively correcting this uneven light degradation is an indispensable step in microscopic image processing and analysis.It has an important impact on subsequent tasks such as quantification,mosaicking,and segmentation.The existing illumination correction models for microscopic images can be broadly categorized into three groups:prospective,retrospective single-image,and retrospective multi-image.Prospective methods,which rely on the reference image obtained simultaneously in the microscopic imaging process to achieve illumination correction,require extra effort from the operator.Conversely,retrospective methods have been widely studied for their convenience and practicality without the need for additional imaging operations.However,the existing retrospective methods focus on the illumination correction of single-channel images.Even the state-of-the-art multi-image approaches based on the statistical properties of image sequence have problems when used in multi-channel images,such as for the color pathological image.In view of this phenomenon,effective retrospective illumination correction algorithms are proposed based on the color microscopic image in this thesis.Main contributions are summarized as follows:1.A single-image correction algorithm is proposed for vignetting that the most common illumination degradation,which is suitable for color microscopic images.According to the radial attenuation prior of vignetting,the problem of uneven illumination of microscopic images is transformed into the corresponding problem of low-light enhancement by quadric surface fitting,and the post-processing is carried out by the enhancement algorithm based on Retinex theory.In the process of transformation,aiming at the vignetting of the color microscopic image and its possible color degradation,we propose to process the value channel of HSV color space,and thus the transformation process is decomposed into two subproblems of brightness and color correction respectively,which avoids the phenomenon of color secondary degradation when the traditional single-image algorithms correct the image of uneven illumination.Compared with the classical single-image correction methods,a more obvious visual quality improvement is achieved in the vignetting correction of the pathological cell image by the proposed algorithm according to the idea of transformation and decomposition.2.A multi-image correction algorithm is studied based on deep convolutional neural network for both vignetting and non-vignetting illumination degradation.The inherent defect of multi-image correction algorithms is that they can’t achieve nonuniform illumination correction when the number of microscopic images under the same illumination source is insufficient.Moreover,most advanced multiimage algorithms at present are easy to produce artifacts when applied to the color microscopic image.The proposed algorithm extracts and integrates the illumination features in the form of a multi-module in series,and ensures that the illumination,color,and texture details of the color microscopic image are optimized in the direction of the gold standard through the weighted loss function.The experimental results show that the model trained on the pathological dataset with vignetting overcomes the defect of the state-of-the-art algorithms and has the effect of illumination correction even for a single pathological image under different illumination sources,and the qualitative and quantitative evaluation results are both the best on the test set.The network is also applied to the preprocessing of color whole slide imaging tiles with non-vignetting illumination degradation,which greatly improves the effect of image mosaicking.