Self-supervised defocus blur kernel estimation with symmetry and smoothness constraints

In microscopic imaging systems, the extremely shallow depth of field often causes defocus-induced blurring as a result of slight vertical movements of biological samples. Accurate blur kernel estimation enables clearer image recovery through deconvolution. Traditional methods for blur kernel estimation and image deblurring rely on idealized assumptions and lack robustness to noise. Meanwhile, deep learning approaches require large datasets, labeled training data, and significant computational resources, limiting their applicability. To address these challenges, we propose a self-supervised neural network that incorporates symmetry and smoothness constraints as loss functions for blur kernel estimation and defocus image restoration. This method leverages feature extraction capabilities of neural networks while using handcrafted priors to guide optimization. It requires no data collection, labeling, or pre-training, making it efficient and adaptable while enabling the handling of complex blur kernels. Experimental results show improved kernel estimation accuracy and enhanced image clarity, with potential applications in medical imaging and biological diagnostics.