the Creative Commons Attribution 4.0 License.
the Creative Commons Attribution 4.0 License.
| 22 Apr 2026
Technical note: a Recognition-assisted Camera for Automated Microscopy (RaCAM)
Abstract. Automated microscopy workflows, including image acquisition, processing, and recognition using artificial intelligence (AI) are getting a growing interest from the scientific community in biogeosciences, as more and more research institutes are actively working on building datasets of images to train artificial convolutional neural networks (CNNs) to identify microscopic objects.
Here, we present a new, affordable, AI-assisted, Raspberry Pi-powered camera, with the first, built-in, and fully auto-mated microscopy workflow (including automated image acquisition, processing and recognition) that can fit any microscope equipped with a regular C-mount (or CS-mount) camera thread. This camera is equipped with an integrated Single-Board Computer (Raspberry Pi 5) and high-resolution camera sensor (12.3 mp), attached together using a 3D-printable adaptor. Us-ing a new open-source software (RaCAM user interface), written using the Python language, and freely downloadable too, the camera is capable of performing automated acquisition of field of view images, segmenting each visible object of interest, and identifying them using trained CNN onnx models in a few seconds as part of a whole automated workflow.
The camera is also adapted to on-field tasks such as core description, biostratigraphy or even palaeoenvironmental reconstructions based on microfossils census data or morphometry, as it can operate without the need for a spare computer and run directly on a power bank. Finally, as the RaCAM workflow relies on images directly captured by the camera, applications can also be extended outside of the microscopy and micropaleontology research fields as any picture acquired with this device can virtually be processed by the automated workflow.
- Preprint
(5075 KB) - Metadata XML
- BibTeX
- EndNote
Status: open (extended)
-
RC1: 'Comment on egusphere-2026-2075', Anonymous Referee #1, 02 Jun 2026
reply
-
AC1: 'Reply on RC1', Martin Tetard, 06 Jun 2026
reply
Comment: The manuscript introduces a camera designed for automated microscopy, utilising Raspberry Pi hardware, a 3D-printed case, and free software. This system acquires images, segments the object of interest within the field of view, and identifies it using a convolutional neural network (CNN). The innovative feature of this solution is that all processing occurs on the Raspberry Pi, requiring only a monitor or screen for operation.
Response: Thank you very much to the reviewer for this review. We are very happy that the reviewer took time to follow all instructions to install and use a RaCAM camera. This feedback is extremely valuable to us. Our first comment is that the workflow that the reviewer tested is the “image segmentation workflow” that uses a python interface for image acquisition, then every particle is segmented using an imageJ script, and then each image is recognised using a trained CNN.
Since publication, we also developed an object detection workflow, more user friendly as it only required a onnx model (yolo11n for example) to run inference on whole FOV directly acquired by the camera, in less than a second, and also inference directly on the live preview. This workflow will be refered to as Object detection workflow. We would be very happy if the reviewer was also able to test this second approach, which is much easier to install, faster to run, and work on for people not used to such automated approaches. We should be able to upload the new workflow and instruction soon on the github repository. A video of it can be seen at: https://www.youtube.com/watch?v=xw3MP37GsEo
Comment: This approach offers several advantages, including a low-cost solution accessible to individuals worldwide who need to count and identify biological objects. The low cost stems from the 3D-printable case and the Raspberry Pi itself. Notably, the image resolution from the high-quality camera sensor (4056x3040, 12 MP) exceeds that of many commercial cameras, such as the Leica ICC50 W/E (2592x1944, 5 MP). Additionally, the segmentation and artificial intelligence software tools provide a significant advantage for anyone looking to take their first steps toward automation.
Response: Thank you to the reviewer for this comment. Of course, we camera being low cost, as we said, it cannot perform as well as hi-end camera regarding video acquisition of fast moving object for example, but for regular microscopy, focusing on high resolution capability is usually better than video quality (again, depending on the application).
Comment: I believe this manuscript has the potential to be relevant for scientific investigations within the journal's scope. As I followed the individual steps, I encountered some technical issues. Specifically, I got stuck during the segmentation process: the program displays "processing," but nothing happens because AutoDiato_RaCAMx40.ijm is designed for radiolaria, not for pollen. I also attempted to download the radiolarian dataset from Carlsson et al. (2023), but the images were already segmented, preventing me from testing the segmentation and identification.
Reponse: Thanks you very much for showing us this mistake. Actually, the file that you are supposed to copy is not AutoDiato_RaCAMx40.ijm but AutoRadio_RaCAMx10.ijm (L172), we made an error here. The fact that nothing was happening in ImageJ is probably because the script path was thus wrong. In order to help reviewers and scientists building their own camera, we will integrate a complete example with already acquired FOV of radiolarian images, so people can train using the workflow with the provided CNN and provided ImageJ script for segmentation. We will also now provide this CNN separately as it weights 97 out of the 98 mo of the RaCAM_software.zip, to prevent having to download it again everytime we update the software.
Again, the provided images, imageJ script for segmentation, and trained CNN (for Eocene radiolarian) should be regarding as an example of capability for the camera, to show how the workflow works, and also for people to use these as a template / base to adapt it for their group of interest. As our team has been working on images of Radiolaria, Pollen, Foraminifera and diatoms, we could also add these imagej segmentation script to the “RaCAM_software.zip” to help people getting started.
If this issue if still encountered, we propose the referee to contact us directly so we can more easily figure out the issue.
Comment: This issue could arise for anyone, so I would like to request some sample images of Radiolaria that cover the entire field of view and work with the provided codes. Additionally, it would be helpful to include examples of other proxies to support the strong statements in the last sentence of the abstract. This would demonstrate that your tool can handle "any micropaleontological image," not just radiolaria.
Reponse: Indeed, this is exactly why you trying to use our tutorial and finding error is a very precious feedback for us as, we did it from scratch several times, but still missed such error in the tutorial. For the segmentation workflow, models are heavy and cannot be easily stored in our github (about 100mo each). However, we remind here that models are “fixed” and won’t perform better or worst on the RaCAM than on any computer running ParticleTrieur, which is used to train them. Our research group has trained models for various groups over the past year (Pollen, Diatoms, Radiolaria, Chironomids) and we provide one that was already published (Eocene Radiolaria) as an example here. For other groups models, we refer the reader to original publication (Such as Tetard et al., 2020 for Modern Radiolaria; Bourel et al., 2020 for Pollens; Another pollen CNN will be published soon for another study, another foraminifera CNN will be available soon (publication submitted), and a chironomid model will be submitted soon associated with another publication). Again, our present manuscript attempt to demonstrate the use of a new tools, what are the possibilities and how to use it, but we cannot provide numerous models for each microfossil groups.
Comment: The sentence at line 295 is correct, but it comes across as somewhat strong. The development of the segmentation macro and the ONNX model must take place before fully utilising RACAM. Additionally, the phrase “user-friendly interface” at line 295 seems a bit misleading, as the requirement to close the preview window in order to capture an image, as mentioned at line 235, is somewhat impractical.
Reponse: Thank you, we have edited this sentence. Indeed, closing the preview window is mandatory and operate as a separate process that we could not overcome in our development as attempting to capture an image while the preview is operating would return a “busy” argument.
Comment:
- Based on the RaCAM output shown in Figure 5, you could create a RaCAM input schema that would be applicable for any proxy. This schema can include the following components:
- Hardware: List the items you need to purchase or print (excluding the monitor and keyboard). Please note that you should specify HDMI micro cables, not mini, as mentioned on line 110. Additionally, include M2 screws and their respective lengths.
- Software: This consists of three parts:
- Your RaCAM_software.zip file
- A segmentation macro from ImageJ
- A CNN ONNX model, which must be trained beforehand using the ParticleTrieur program. Alternatively, this model can be created following the last part of your GitHub instructions, which appears below the program screenshot. Please note that this part is not clearly described in the manuscript.
Response: The micro vs mini hdmi mistake has been corrected. We thank the referee for this very good suggestion, we will prepare a figure summarising all hardware and software component, hopefully this will make the whole procedure easier to understand. Regarding the comment about training a model using the last part of our GitHub instructions: Actually, this last part gives instructions on how to use an existing trained model. We did not detail the model training is this manuscript, as it aims to present the Camera and how we put the hardware and software piece together, and as the model training part was already detailed in Marchant et al., 2020, and at https://particle-classification.readthedocs.io/en/latest/ We have edited our instruction to explain this.
Comment: Add to your GitHub instructions that, after installing Miniconda 3, it is necessary to close the window to initialise conda
Reponse: We added it, thanks.
Comment: Program Behaviour: The requirement to close the preview window in order to capture an image is a significant issue, especially since there is no "X" button in the corner to close it.
Reponse: The preview window should have a x button to close it. As visible in fig.5. We recently updated the camera and found that the preview window would open in the top left corner with no visible button bar. An error likely due to a recent system update. To sort it out, you can click on Alt or the Windows/Raspberry key on your keyboard, and you should be able to drag the preview window out of the corner by click and drag it with the left mouse button (or click on alt+f4 to close it). There might be some compability issue with recent updates. To prevent this we have updated the RaCAM software to force the preview in a window that can be moved around and will appear in a specific position. We will upload it shortly.
Comment: The JPEG output format contains compression artefacts. It would be beneficial to add another option for saving data, ideally using an image format like OME-TIFF that includes standardised metadata for microscopy.
Reponse: Tests have been conducted in “Tetard, M., Carlsson, V., Meunier, M., and Danelian, T.: Merging databases for CNN image recognition, increasing bias or improving results?, Mar. Micropaleontol., 185, 102296, https://doi.org/10.1016/j.marmicro.2023.102296, 2023”. Showing no difference in CNN accuracy between using the same FOV image saved as TIFF, JPEG, and 90% quality JPEG, so we really encouraged using JPEG with the camera to use files as light as possible for the camera. The possibility to save in tiff format in the RaCAM software can however be changed easily in the code, and we will try to add this possibility in the near future, together with the PNG format. Please let us know if you feel like other option are missing from the interface.
Comment: The magnification drop-down menu should include both objective and projective information; my microscope has an objective of 50x and a projective of 0.5.
Response: The magnification dropdown menu is just used for naming convention of the acquired image, to include it in the metadata of the image. Even if your magnification is not in the dropdown menu, you should be able to manually write it in the field. We will had some common magnification in the dropdown menu.
Comment: Reopening the program resets the settings from the last session – again, not very user-friendly.
Reponse: We thank the referee for this excellent suggestion, I was actually also having a hard time re-entering all the parameters everytime I used the camera on field. We had a button to save most of the parameters (the one that are numerical, as well as the core and sample names); and a button to load it instantly.
Comment: I understand that the program has five functions: Live View, Snapshot (Image Acquisition), Image Processing, Image Recognition, and Census Data. However, only four red buttons are available. What happens if all functions are set to "yes"? I would expect that if one function is set to "yes," the others would automatically be set to "no."
Response: The workflow actually work differently, and we thank the reviewer for this, we will edit the text and software instructions to make it clearer. The live view only show a live preview, no matter if Image Acquisition, Image Processing, and Image Recognition are set to “yes”, or “no”. These parameters only affect what the software will do regarding snapshot and batch processing. If only image acquisition is set to “yes”, it will just take a picture. If image recognition is also set to yes, “snapshot” will take a picture and process it. If image recognition is also set to yes, “snapshot” will take an image, segment it (processing) and identify every particle segmented. If image acquisition is set to “no” you can’t do a “snapshot”, but you can “batch process” a core folder, containing sample subfolders, containing FOV images, either by only doing image processing, only image recognition, or both image processing and recognition. We will update the instructions to make it clearer.
Thanks again for all the constructive comments.
Citation: https://doi.org/10.5194/egusphere-2026-2075-AC1 - Based on the RaCAM output shown in Figure 5, you could create a RaCAM input schema that would be applicable for any proxy. This schema can include the following components:
-
AC1: 'Reply on RC1', Martin Tetard, 06 Jun 2026
reply
Viewed
| HTML | XML | Total | BibTeX | EndNote | |
|---|---|---|---|---|---|
| 355 | 78 | 20 | 453 | 18 | 20 |
- HTML: 355
- PDF: 78
- XML: 20
- Total: 453
- BibTeX: 18
- EndNote: 20
Viewed (geographical distribution)
| Country | # | Views | % |
|---|
| Total: | 0 |
| HTML: | 0 |
| PDF: | 0 |
| XML: | 0 |
- 1
The manuscript introduces a camera designed for automated microscopy, utilising Raspberry Pi hardware, a 3D-printed case, and free software. This system acquires images, segments the object of interest within the field of view, and identifies it using a convolutional neural network (CNN). The innovative feature of this solution is that all processing occurs on the Raspberry Pi, requiring only a monitor or screen for operation.
This approach offers several advantages, including a low-cost solution accessible to individuals worldwide who need to count and identify biological objects. The low cost stems from the 3D-printable case and the Raspberry Pi itself. Notably, the image resolution from the high-quality camera sensor (4056x3040, 12 MP) exceeds that of many commercial cameras, such as the Leica ICC50 W/E (2592x1944, 5 MP). Additionally, the segmentation and artificial intelligence software tools provide a significant advantage for anyone looking to take their first steps toward automation.
I believe this manuscript has the potential to be relevant for scientific investigations within the journal's scope. As I followed the individual steps, I encountered some technical issues. Specifically, I got stuck during the segmentation process: the program displays "processing," but nothing happens because AutoDiato_RaCAMx40.ijm is designed for radiolaria, not for pollen. I also attempted to download the radiolarian dataset from Carlsson et al. (2023), but the images were already segmented, preventing me from testing the segmentation and identification.
This issue could arise for anyone, so I would like to request some sample images of Radiolaria that cover the entire field of view and work with the provided codes. Additionally, it would be helpful to include examples of other proxies to support the strong statements in the last sentence of the abstract. This would demonstrate that your tool can handle "any micropaleontological image," not just radiolaria.
- Your RaCAM_software.zip file
- A segmentation macro from ImageJ
- A CNN ONNX model, which must be trained beforehand using the ParticleTrieur program. Alternatively, this model can be created following the last part of your GitHub instructions, which appears below the program screenshot. Please note that this part is not clearly described in the manuscript.
Carlsson V., Danelian T., Tetard M., Meunier M., Boulet P., Devienne P., & Ventalon S. (2023): Convolutional neural network application on a new middle Eocene radiolarian dataset. – Marine Micropaleontology.