Research On Autofocusing Algorithms Based On Whole Slide Imaging

As an effective mean to observe the microscopic world,optical microscope has been widely used in biomedical and scientific research.In order to capture a clear image,the imaging system must adjust the position of the sample to the focal plane.We call this process focusing.In the traditional microscope,focusing is completed manually which is not efficient.With the development of the imaging system,combining with image processing technology,a new technology called Whole Side Imaging i.e.WSI,appears.This technique can automatically complete the image focusing,image acquisition and image stitching which is very promising.The core of WSI is the autofocusing algorithm which completes the focusing process automatically and makes the final image clear.The performance and speed of the algorithm affect the imaging quality and efficiency of WSI system.The current mainstream autofocusing algorithms include the Depth From Focus(DFF)method and the Depth From Defocus(DFD)method.The DFF method needs to acquire a dozen images of the same target which usually takes a long time.The DFD method only needs one or a few images,but it relies on an accurate defocus degradation model which cannot be accurately obtained.So the focusing error of DFD is relatively big.In view of the shortcomings of the existing autofocusing algorithm,we proposed two new algorithms related to autofocusing:1)Based on the asymmetry of positive and negative defocusing in the optical imaging system,a deep cascade autofocusing network is proposed.The input of the algorithm is defocusing blurred image,and it's output is the distance between the current image position and the clear imaging position,which called defocus distance.Utilizing this difference between positive and negative images and learning two corresponding regression networks can improve performance of the autofocusing algorithm.To improve the generalization,channel attention mechanism is used in the model.Moreover,channel normalization strategy and the data augmentation strategy based on color channel exchange are added in the training stage.Experiments prove that our method has smaller focusing error and stronger generalization ability compared with the current best single image autofocusing algorithm.And our algorithm is faster than traditional methods,2)Based on the deep cascade autofocusing network proposed in this paper,an image deblurring network based on pre-focusing strategy is proposed.The algorithm's input is blur image,and the output is clear image.In a larger sense,this deblurring algorithm is also an autofocusing algorithm.The algorithm first uses the pre-focusing strategy to limit the defocusing distance of the sample,so as to reduce the difficulty of deblurring.In order to improve the performance of deblur algorithm,channel attention mechanism,residual learning strategy and a weight loss based on defocus distance are used during the training process.Experimental results show that the performance of proposed algorithm is superior to the existing advanced deblurring algorithms,and it can directly produce clear imaging from the defocus blurred image without reducing the imaging quality.This algorithm don't have to move the mechanical device,and the imaging speed is fast.So it can handle a large number of sample quickly.This algorithm is suitable for the WSI systems which require high imaging speed.