Search the Community

Hello everyone, I have quite a few projects going on and I'm going to be a bit more active on here to share things. Sorry I've been a bit quiet! I see there are a few threads on here about CT scanning, 3D printing and segmentation, but I thought I would add this one. I've been doing it for a few years, but just bit the bullet and bought a machine custom built to do this. I really think we should create a sub section of fossil preparation for CT scanning and segmentation? Resources are scattered around the internet and it would be great to document them all on here. I have lots to learn and I'd like to "upload" this knowledge here as I aquire it. E.g. Hardware set up, software, things like digitization tips and tricks (stylus pen and tablet vs mouse). The same sort of thing we have for physical prep on here: set-up, tools and techniques. Anyway. I just acquired a very powerful computer and I fired it up last night. I have a CT scan of a Pliocene gannet skull that was found last year that I've been sitting on. At least I think it is gannet. I am using imageJ to do the segmentation. Segmentation is just the process of telling the computer what is bone, what is rock and what is air. I don't have a photo of the concretion, but here is a 3D model above. It is about 10cm long. This is looking from above. Here is one of 760 slices from the CT scan of the skull. A vertical slice with the top of the skull at the top of the image. Pixel size is about 30 microns! Here is the view after ONE round of training the computer. I selected some areas of bone, some of rock and some of air. Then the computer thought about it, using 160 GB of ram (out of my total 192 GB) and the latest Intel chipset in a water-cooled CPU to classify every pixel as bone (red), rock (green) or air (purple). It does this for every of the 760 slices. This is a first pass. You can go back and train the computer further and correct it. It gets better with each round. Here is the first reconstruction of the skull. You can see there is still a bit of noise. I could get rid of that with a few more learning phases. A lot of loose pixels could be removed in rendering software such as Blender too. Hope you enjoyed this. I'll keep you posted as I improve the model. And I'd like to 3D print it at the end!

(totally new to this forum, let me know if there's a more appropriate place to post this!) Hey all! I found a video showcasing a fossil hunting method that I'm super fascinated by! It's a drone that uses the fact that many fossils fluoresce under UV light to fly around and find fossils autonomously. This seems like an absolutely fantastic way to hunt for fossils over a large area. I have experience with drones and electronics and I'm interested in making a similar drone, hopefully a bit simplified, maybe open-source and user friendly etc. If this method works well to find fossils, getting this tool into the hands of more researchers would be an awesome way to find many more fossils. So I'm looking more into just how effective this method is; particularly how often it is that a fossil is fluorescent, especially vertebrate fossils. I did find this paper that gives some good info but it is largely focused on smaller vertebrates and only from a single formation. Basically any information you have about fossil fluorescence under UV light I'd love to hear! I'm especially interested in figuring out how broad of a phenomenon this is. Do most vertebrate fossils from most formations fluoresce? Is it only smaller fossils or only some formations? The better this method is for finding fossils, the more effort I want to put into making a simple, reliable drone employing the method. I'm also interested in automating the process with machine learning down the line. I read that fossils have a typical fluorescence wavelength that is different from most other things you might find out there, which gives you a clear signal to look for. The eventual result could be: you let a drone go fly around autonomously for a couple hours, recharging it as needed, and then it automatically gives you a list of GPS coordinates to go check out and the images associated with them. (Also, if anyone wants to collaborate on this project, please let me know, whether your skills are more on the electronics/software or paleontology side!) Disclaimer: I'm very aware that drone flight is not allowed in many typical fossil bed areas, definitely won't ever be flying anything without approval.

Although this paper involves meteorites, that drones were "taught" to recognize meteorites using machine learning suggests that there is the potential that the same can be done for larger fossils such as vertebrate remains. The paper is> Anderson, S., Towner, M., Bland, P., Haikings, C., Volante, W., Sansom, E., Devillepoix, H., Shober, P., Hartig, B., Cupak, M. and Jansen‐Sturgeon, T., 2020. Machine learning for semi‐ automated meteorite recovery. Meteoritics & Planetary Science. First published: 01 December 2020 ARXIV PDF file for those people who lack subscription / library access: https://arxiv.org/pdf/2009.13852.pdf Yours, Paul H>

Hello Fellow Forum-Goers, Lately I have been somewhat inactive on the forum, and also have not had the opportunity to go fossil hunting in New Jersey since I am at college. But those things do not deter me though. I am here today to tell you about a project I have been doing with fossil classification, specifically classification of some fossil species from the Cretaceous of New Jersey. The goal is to be able to give my computer of a fossil and have it tell me with a certain degree of confidence the probability that it is any one of several New Jersey Cretaceous fossil species. For this project, I began by taking photos of some of my fossils. Here are some examples of the what the photos looked like: Anomoeodus phaseolus Ischyodus bifurcatus Brachyrhizodus wichitaensis The data consisted of around 150 photos, spanning across 6 fossil species. Not represented in the photos above were: Belemnitella americana Enchodus petrosus Ischyrhiza mira To later label this data, I wrote a csv file with labels. From the contents of this file, you can see how many of each specie there were. Note how the common name for these species are used as the labels. id,species IMG_4749,Belemnite-1 IMG_4780,Belemnite-2 IMG_4812,Ray-1 IMG_4813,Ray-2 IMG_4814,Ray-3 IMG_4815,Ray-4 IMG_4816,Ray-5 IMG_4817,Ray-6 IMG_4818,Ray-7 IMG_4819,Ray-8 IMG_4820,Ray-9 IMG_4821,Ray-10 IMG_4822,Ray-11 IMG_4823,Ray-12 IMG_4824,Ray-13 IMG_4825,Ray-14 IMG_4826,Ray-15 IMG_4827,Ray-16 IMG_4828,Ray-17 IMG_4829,Ray-18 IMG_4830,Ray-19 IMG_4831,Ray-20 IMG_4832,Ray-21 IMG_4833,Ray-22 IMG_4834,Ratfish-1 IMG_4835,Ratfish-2 IMG_4836,Ratfish-3 IMG_4837,Ratfish-4 IMG_4838,Ratfish-5 IMG_4839,Ratfish-6 IMG_4840,Ratfish-7 IMG_4841,Ratfish-8 IMG_4842,Ratfish-9 IMG_4843,Ratfish-10 IMG_4844,Ratfish-11 IMG_4845,Ratfish-12 IMG_4846,Ratfish-13 IMG_4847,Ratfish-14 IMG_4848,Ratfish-15 IMG_4849,Ratfish-16 IMG_4850,Ratfish-17 IMG_4851,Ratfish-18 IMG_4852,Ratfish-19 IMG_4853,Ratfish-20 IMG_4854,Ratfish-21 IMG_4855,Ratfish-22 IMG_4856,Ratfish-23 IMG_4857,Ratfish-24 IMG_4858,Ratfish-25 IMG_4859,Ratfish-26 IMG_4860,Ratfish-27 IMG_4861,Ratfish-28 IMG_4862,Ratfish-29 IMG_4863,Ratfish-30 IMG_4864,Ratfish-31 IMG_4865,Ratfish-32 IMG_4866,Ratfish-33 IMG_4867,Ratfish-34 IMG_4868,Enchodus-1 IMG_4869,Enchodus-2 IMG_4870,Enchodus-3 IMG_4871,Enchodus-4 IMG_4872,Enchodus-5 IMG_4873,Enchodus-6 IMG_4875,Enchodus-7 IMG_4876,Enchodus-8 IMG_4877,Enchodus-9 IMG_4878,Enchodus-10 IMG_4879,Enchodus-11 IMG_4888,Enchodus-12 IMG_4889,Enchodus-13 IMG_4890,Enchodus-14 IMG_4891,Enchodus-15 IMG_4892,Enchodus-16 IMG_4903,Pychodont-1 IMG_4905,Pychodont-2 IMG_4906,Pychodont-3 IMG_4907,Pychodont-4 IMG_4908,Pychodont-5 IMG_4909,Pychodont-6 IMG_4910,Pychodont-7 IMG_4911,Pychodont-8 Now, with the labels and data. I began to make a program that fed in the images and then used Keras ( a machine learning library that has the tools for something called a convolutional neural network) in the programming language Python. Here is the beginning of the code: import numpy <- This gets me NumPy, which allows for easy use of vectors and matrices to work with data import collections <- This allows me to make better data structures called "dictionaries" import os <- This allows me to get the path of the image in my computer import imageio <- This allows me to write edited images to other folders from PIL import Image <- This allows me to manipulate the images, in way such as flipping or rotating. from random import shuffle <- This allows me to randomly shuffle the training data. This code puts each fossil image and its label into something like a container together, this "container" is called a dictionary species_dictionary = collections.OrderedDict() our_file = open("fossil_labels.csv","r") file_contents = our_file.read() file_contents = file_contents.split('\n') for iteration in range(1,len(file_contents)): file_contents[iteration] = file_contents[iteration].split(',') species_dictionary[file_contents[iteration][0]] = file_contents[iteration][1] I will skip the other code and now discuss convolutional neural networks, which are used in image classification So, right now I have all the fossil image with their label. The network can now use these to find clusters of pixels in a given photos that correspond to a particular fossil species. Overtime, the network letters that this or that cluster of pixels is common to one a single fossil species. Then, it can recognize that cluster in a new or novel image that it has not been trained on. Here are the convolutional layers: model = Sequential() model.add(Conv2D(32, kernel_size = (3, 3), activation='relu', input_shape=(IMG_SIZE, IMG_SIZE, 1))) model.add(MaxPooling2D(pool_size=(2,2))) model.add(BatchNormalization()) model.add(Conv2D(64, kernel_size=(3,3), activation='relu')) model.add(MaxPooling2D(pool_size=(2,2))) model.add(BatchNormalization()) model.add(Conv2D(64, kernel_size=(3,3), activation='relu')) model.add(MaxPooling2D(pool_size=(2,2))) model.add(BatchNormalization()) model.add(Conv2D(96, kernel_size=(3,3), activation='relu')) model.add(MaxPooling2D(pool_size=(2,2))) model.add(BatchNormalization()) model.add(Conv2D(32, kernel_size=(3,3), activation='relu')) model.add(MaxPooling2D(pool_size=(2,2))) model.add(BatchNormalization()) model.add(Dropout(0.2)) model.add(Flatten()) model.add(Dense(128, activation='relu')) model.add(Dense(5, activation = 'softmax')) model.compile(loss='binary_crossentropy', optimizer='adam', metrics = ['accuracy']) model.fit(training_images, training_labels, batch_size = 50, epochs = 10, verbose = 1) I do not expect this code to be fully understood.The network uses weights or sensitivities to different pixel clusters. Then as it learns how its predictions for a photo compares to the actual training photo I gave it, it updates the weights to reflect this. By the end this "error loss" should reach towards 0, and when it does, we know that its predictions correspond very close with the actual photo, now allowing it to classify fossil images for these six species well. If it were training on 10 species it would classify all 10 well. Here is the output of the training: # Epoch 1/10 # 164/164 [==============================] - 497s 3s/step - loss: 0.3975 - acc: 0.8329 # Epoch 2/10 # 164/164 [==============================] - 139s 846ms/step - loss: 0.1026 - acc: 0.9610 # Epoch 3/10 # 164/164 [==============================] - 139s 848ms/step - loss: 0.0427 - acc: 0.9902 # Epoch 4/10 # 164/164 [==============================] - 126s 771ms/step - loss: 0.0232 - acc: 0.9939 # Epoch 5/10 # 164/164 [==============================] - 119s 728ms/step - loss: 0.0153 - acc: 0.9963 # Epoch 6/10 # 164/164 [==============================] - 2258s 14s/step - loss: 0.0066 - acc: 0.9976 # Epoch 7/10 # 164/164 [==============================] - 141s 861ms/step - loss: 0.0047 - acc: 1.0000 # Epoch 8/10 # 164/164 [==============================] - 135s 824ms/step - loss: 0.0048 - acc: 1.0000 # Epoch 9/10 # 164/164 [==============================] - 132s 803ms/step - loss: 0.0027 - acc: 1.0000 # Epoch 10/10 # 164/164 [==============================] - 122s 746ms/step - loss: 0.0043 - acc: 1.0000 You can see that the loss keeps going down with more and more training. For the future, I definitely need to take more photos to get more data and allow it to train on a graphical processing unit (GPU) as opposed to the normal CPU that you use on a laptop. The GPU is better at parallel processing and can train the network in seconds (on my computer it took 15 minutes). Well that is the current state of the project. I still need to do more but thank you for staying here till the end. I hope you have a nice day. -Trevor