Figure 2. Deep-learning-based feature recognition for autonomous grid navigation.

Sample images show the performance of the square (A) and hole (B) detectors applied to gold (left) and carbon (right) grids. (A) Automatic detection of squares and classification into six different classes: small, cracked, dry, contaminated, good, and partial (white scale bars are 100 μm). Representative examples of squares assigned to each class and corresponding detection precision values are shown (bottom panel). (B) Hole detection performance on representative square images extracted from gold and carbon grids. The hole detector implements a classification step to filter out contaminants (shown in yellow) and increases hole detection precision (shown as pink circles) (white scale bars are 10 μm).

Figure 2—figure supplement 1. Object detection training strategy.

Atlas (A) and Windows (B) where manually picked and annotated. The annotated dataset was augmented using a combination of rotations, translations, contrast variations, magnifications, and flip transforms. The models were trained against their specific architectures.