Deblurring microscopic image by integrated convolutional neural network

Abstract

In microscopic vision, the sharpness of edge profile seriously affects the accuracy of measurement. However, various factors in complex environments often blur the edge profile. Herein, a CNN-related deblurring algorithm is proposed to reconstruct the real edge features from the blurred microscopic images. Based on the convolutional neural network, a network called I&R Net is built up, in which both the inception block and the residual channel attention block are integrated. Moreover, the loss functions are constructed with the additional factor of gradient by Sobel operator and Laplacian operator to enhance the inductive bias of I&R Net on real size. The results show that the I&R Net improves both the general image quality and the accuracy. Additionally, the I&R Net presents relatively strong resistance to the influence of feature-rotating.