In nanoparticle imaging, particle clusters and large impurity particles in the defocused position cause bright spots, thus hindering the existing focusing algorithms in realizing the autofocus function. This study used binarization segmentation based on the Otsu algorithm, as well as morphological opening and closing methods, to aggregate the dispersed diffuse spots into one area. Furthermore, the connected domain labeling method was used to filter out large regions of the spot area. A four-neighborhood level-diagonal square function and threshold-four-neighborhood level-diagonal square root function were constructed and used as the evaluation indicators for the coarse and fine focus, respectively, thereby improving the accuracy and reliability of autofocus search. The defocus sequence diagram was obtained and the proposed algorithm was compared to the five commonly used evaluation algorithms. The results demonstrate that the proposed autofocus evaluation algorithm is highly robust, unbiased, and unimodal.