In response to the challenge where current focus evaluation methods struggle to accurately assess image focus amidst fluctuations in image scene content,a focus evaluation method,fully equivalent to Gaussian blur standard deviation,is introduced based on the"classification+fitting"concept.Firstly,a dataset for image focus classification,annotated with finite Gaussian blur standard deviations,is established.Subsequently,an asymmetric kernel convolution neural network(AKC-net)is developed to derive image Gaussian blur standard deviation classification scores.Finally,a cubic spline interpolation function is applied to fit the classification score of the AKC-net fully connected layer output and the corresponding Gaussian blur standard deviation,using the standard deviation linked to the maximum score as the focus evaluation for the images.Simulation experiments and real shooting experiment are carried out on the Waterloo dataset and real shooting images respectively.The experimental findings indicate that the proposed method achieves an average classification accuracy of up to 97.7％across various focus images.The root mean square error and mean absolute error between the obtained evaluation results and the true Gaussian blur standard deviation values are less than 0.07.Furthermore,the focus measurement values derived from real shooting images are independent of the image content,enabling the absolute evaluation of the image focus.
degree of focus evaluationGaussian blur standard deviationasymmetric kernel convolution neural networkcubic spline interpolation
NETL
NSTL
万方数据
