Deep Learning Analysis of Swipe Blood Stains to Infer Fluid-Surface Interactions

Blood stains created by blood-coated objects swiped along non-absorbing surfaces, commonly found at crime scenes, reveal underlying squeeze flow dynamics and the subsequent drying of thin films. This study employs deep learning to classify swipe stains based on both the type of blood-coated object and the direction of the swipe. A growing dataset of 570 swipe images, generated using 12 different objects, was augmented through image flipping and rotation. Previous machine learning analyses of stain images have demonstrated potential for determining fluid-surface interactions during drying and staining processes [1]. In the present study, the images were processed using an open-source convolutional neural network to extract features, followed by a custom multi-output neural network for classification. For objects included in the training set, the average error rates were approximately 5% for object type prediction and 10% for swipe direction. When tested on previously unseen objects, the average error rate for direction classification remained below 20%. These results suggest that dried swipe patterns encode rich information about the object's surface geometry and motion, motivating further investigation into their underlying fluid dynamics. Two deep learning strategies are presented to identify discriminative flow features. Canonical microscale flow scenarios are also discussed to guide experimental design across the broad parameter space governing squeeze flow and the drying behavior of blood swipe patterns.

[1] Kim, Namwon, Zhenguo Li, Cedric Hurth, Frederic Zenhausern, Shih-Fu Chang, and Daniel Attinger. “Identification of Fluid and Substrate Chemistry Based on Automatic Pattern Recognition of Stains.” Analytical Methods 4, no. 1 (2012): 50–57. https://doi.org/10.1039/c1ay05338h.