maXerial

Left: Original. Right: Corrected and denoised with our latest model.  The video is for marketing. The stats are real:  - 5x more voids detected - 10x more void content measured - Largest void 3x bigger than initial reconstruction suggested - 20 false-positive voids eliminated Same data, same machine, just better algorithms.  More confidence in your results, higher quality with the same effort. Reach out if you want to learn more: info@maxerial.io maXerial #IndustrialCT #VoidAnalysis #SMT

These images are ugly. Spending 8 hours a day with them is worse. Magnetic tiles are the curved ferrite arcs inside DC motors. They are everywhere: in your car, in your drill, in your kitchen mixer. Each one gets inspected for surface defects before it ships. The defects are dark gray marks on a dark gray ceramic surface, sub-millimeter, under industrial lighting. Not what you want to stare at for hours. This is where AI can really help. We trained a binary U-Net on 115 images from Kaggle to a validation IoU of 0.832. For training, we used our 1-Click-AI-Trainer. It took only minutes. No code, no cloud, just a few clicks. Some jobs are better done by machines. Dataset: Huang, Qiu, Yuan, Surface Defect Saliency of Magnetic Tile (2018). https://sl1nk.com/e29azql #1ClickAITrainer #IndustrialAI #DefectDetection #MachineVision #maXerial

Ball screws steer your car and land your plane. When pitting starts on the raceway, the part is on borrowed time. Not what you want in a steering rack or a plane wing. Pitting damage is hard to spot before it propagates and takes expert knowledge to identify reliably. So we trained an AI model to find it. The dataset [1] from Schlagenhauf et al. at KIT consists of 394 annotated images with varying illumination and contrast. Defects range from a few pixels to larger areas, as visible in the video. Training details: U-Net with MobileNetV2 backbone, 1024² inputs, focal loss, Adam optimizer, 80/10/10 split, 200 epochs. 80 minutes of training on two RTX 3060s. IoU 0.9 on both training and validation splits. Total work on the human side: upload, select model, click train, click predict. 1 to 2 minutes. That's the 1-Click-AI-Trainer. If you have parts and inspection data sitting in a folder somewhere, we're curious what it could find. Send them to us for a first assessment → info@maxerial.io [1] BSData by Schlagenhauf, Landwehr & Fleischer (KIT), CC-BY-SA-4.0 Repository: github.com/2Obe/BSData Paper: Data in Brief 39 (2021) 107643, arxiv.org/abs/2103.13003 #IndustrialAI #MachineVision #DefectDetection #BallScrew

The flicker is hypnotic: 30 seconds inside a neural network, looking at an X-ray of a BGA. It is a convolutional neural network. What you are watching is the same input image passing through its learned filters, layer by layer, channel by channel. Edges and gradients at shallow depth, shapes and textures further in, object-level responses near the top. That is why it is called deep learning. None of it hand-crafted. The network built all of it from annotated examples during training. The last 5 seconds are the output: every solder joint located, scored around 0.95 confidence. This is a model we recently fine-tuned for BGAs. Trained from a small set of annotated X-rays with our 1-Click-AI-Trainer. Next part, next customer, same workflow. #1ClickAITrainer #BGAInspection #IndustrialAI #NDT #maXerial

#Easter egg hunt, maXerial style. We used a semantic segmentation model to identify football players on the field because, apparently, that's how we celebrate holidays now. The setup: 512 CC0-licensed images from Kaggle (80/10/10 split), a UNet with MobileNet backbone, Adam optimizer with cosine annealing warmup, SparseCategoricalCrossentropy loss, 1024 × 1024 px inference. Training time: 35 minutes, 150 epochs, using our #1ClickAITrainer, running inference on the maXerial #AIEdgeKit, completely offline. But the thing is: You don't need to know any of this: No cloud. No PhD required. Label → train → deploy. Happy Easter from the team. May your loss curves be smooth and your false positives be few. Links: 1-Click-AI-Trainer: https://lnkd.in/eRVMHVAQ Kaggle Dataset: Football Player Segmentation (CC0 1.0), https://lnkd.in/eJxazfyp #MachineLearning #ComputerVision #EdgeAI #SemanticSegmentation #maXerial

#hiring #stellefrei