Robustness of machine learning to color, size change, normalization, and image enhancement on micrograph datasets with large sample differences

Appropriate image preprocessing could improve machine learning performance, but the robustness of machine learning to preprocessing methods in micrograph datasets with significant sample differences remains unexplored. Here we collected hundreds of optical micrographs varied in color, contrast, size and brightness and tensile strength from published literature. After preprocessing including color transformation, size adjustment, normalization and image enhancement, classification model of whether it was texture or not and tensile strength prediction model were established using transfer learning and VGG16 model. Our results showed comparable accuracy between classification models using grayscale and color micrographs. Based on grayscale images, classification model utilizing center cropping achieved an average accuracy 1.09% higher than model employing scaling, while coefficient of determination (R2) of regression model showed an average increase of 0.0404. The combination of cropping and dividing by 255 was preferable for classification model, yielding an accuracy of 90.79%. Similarly, the combination of cropping and min–max normalization had a better effect on regression model, yielding R2 of 0.2167. Histogram equalization emerged as the optimal technique for improving classification model, yielding 2.63% increase in accuracy compared to models without image enhancement. For regression models, gamma correction exhibited the highest improvement effect, with an R2 of 0.2681.