1:45 PM - 3:15 PM
[SCG55-P06] Applicability of object detection to microfossil research: Implications from deep learning models to detect ichthyoliths using YOLO-v7
Keywords:machine learning, deep learning, object detectioin, seafloor sediments, microfossils, ichthyolith
Microfossils have constrained depositional ages and environments of various kinds of seafloor sediments, as well as providing high resolution, detailed records of evolutionary processes. Although observational morphological information of fossils provides key constraints to their identification, the traditional observation method relied on manual work. Automation of the observation processes both saves time and also provides opportunities for new discoveries by increasing the number of fossils that can be observed.
Recently, computer vision technologies are developing rapidly. In particular, image processing using deep learning has been applied to various fields including Earth science [e.g. 1]. Although deep learning has come to be applied to observation of microfossils from images [e.g. 2, 3, 4], most of the previous studies applied a technique called image classification, which requires individual pictures for all the particles placed on the glass slide. To achieve this, previous studies proposed following three steps; (1) to place the particles separately on slide, (2) to take images of the whole part of the slide and (3) to recognize outlines of each particle by thresholding using the brightness of images. However, this process still requires significant manual effort to place the particles separately, and it is difficult to apply traditional thresholding to transparent particles.
We have proposed another deep learning technique called object detection, which can predict the position and classes of multiple objects in an image at the same time [5]. Here, we applied one of the latest object detection models “YOLO-v7” [6] to detect microfossils of fish teeth and denticles (ichthyoliths), and found that object detection is applicable to detecting and counting the fossils in multiple classes. Since our method can detect fossils that are considered to be difficult to detect in the previous methods, such as those overlapping with other particles or having similar brightness with background, we believe that the object detection would provide efficient observation to various microfossils.
References : [1] Mimura et al. (2022) TechRxiv. [2] Itaki et al. (2020) Progress in Earth and Planetary Science, 7 1-7. [3] Tetard et al. (2020) Climate of the Past, 16, 2415-2429. [4] Richmond et al. (2022) Geochemistry, Geophysics, Geosystems, e2022GC010689. [5] Mimura et al. (2022) Applied Computing and Geosciences, 16, 100092. [6] Wang et al. (2022), arXiv.
Recently, computer vision technologies are developing rapidly. In particular, image processing using deep learning has been applied to various fields including Earth science [e.g. 1]. Although deep learning has come to be applied to observation of microfossils from images [e.g. 2, 3, 4], most of the previous studies applied a technique called image classification, which requires individual pictures for all the particles placed on the glass slide. To achieve this, previous studies proposed following three steps; (1) to place the particles separately on slide, (2) to take images of the whole part of the slide and (3) to recognize outlines of each particle by thresholding using the brightness of images. However, this process still requires significant manual effort to place the particles separately, and it is difficult to apply traditional thresholding to transparent particles.
We have proposed another deep learning technique called object detection, which can predict the position and classes of multiple objects in an image at the same time [5]. Here, we applied one of the latest object detection models “YOLO-v7” [6] to detect microfossils of fish teeth and denticles (ichthyoliths), and found that object detection is applicable to detecting and counting the fossils in multiple classes. Since our method can detect fossils that are considered to be difficult to detect in the previous methods, such as those overlapping with other particles or having similar brightness with background, we believe that the object detection would provide efficient observation to various microfossils.
References : [1] Mimura et al. (2022) TechRxiv. [2] Itaki et al. (2020) Progress in Earth and Planetary Science, 7 1-7. [3] Tetard et al. (2020) Climate of the Past, 16, 2415-2429. [4] Richmond et al. (2022) Geochemistry, Geophysics, Geosystems, e2022GC010689. [5] Mimura et al. (2022) Applied Computing and Geosciences, 16, 100092. [6] Wang et al. (2022), arXiv.