Fossil Bone Colour Identification Through a Combined Approach of Deep Learning and Machine Learning

In bones, color changes can indicate a variety of taphonomic events, such as burial speed and exposure to environmental factors. The most useful information is often the color change within a single bone, which can reveal local environmental conditions. Usually, the Munsell color system is used for this task, however, Munsell colors can be difficult to differentiate due to the inherent subjectivity of user perception, making the process time-consuming and error-prone. To improve consistency and objectivity in color analysis using the Munsell system, we propose automatic bone element recognition followed by color matching with Munsell colors through unsupervised clustering. The procedure involves two steps. First, bones in the image are extracted by a trained Mask R-CNN, generating isolated images of individual bones. Then, the pixels of each bone are extracted and converted to SRGB, CIELAB, or CIELAB with CIEDE2000 correction color spaces. For clustering, we applied K-means or Gaussian Mixture Models to all color spaces. The resulting color centroids are compared to pre-converted Munsell chip annotations in the same color space, and an average color difference (ΔE) is computed. We analyzed eight images and identified eleven bone elements using Mask R-CNN. The color analysis yielded a range of ΔE values for each image. The exception is the K-means clustering in the CIEDE2000 space setup, where we got an acceptable range, with values varying from 0.53 (unnoticeable difference) to 5.65 (near the acceptable threshold). The results indicate high variance under the current setup (except for K-means with CIEDE2000), and other perceptual color spaces or clustering algorithms may further improve performance. Nevertheless, our findings demonstrate that matching bone colors from uncontrolled photographic sources is feasible and can be achieved with satisfactory accuracy, aiding specialists in faster and more reliable decision-making. [CAPES 88887.683402/2022-00 e  CNPq 169842/20 e CNPq 305576/2021-6 e FAPERJ E-26/204.464/2024].