Artificial Intelligence: Defective Product Inspection/Detection AI Sensor using Image Recognition

We have started offering the unsupervised anomaly detection of the AI visual inspection platform "DeepAge Sensor" as an inference library (DLL for Windows) that can be integrated into inspection devices. The model, trained only on good product images, can be executed on the device without the need for defective product images in the training process. It operates solely on the CPU without requiring GPU or CUDA, and it completes all processes internally without external communication. It can be integrated using a C language interface from C/C++/C#, and it can be shipped with our company's code signing (Authenticode).

We have added a multi-stage fine-tuning feature to the AI visual inspection platform "DeepAge Sensor." You can specify the model created through fine-tuning as the base model for the next additional learning. By simply uploading additional images collected on-site, you can update the model without having to retrain it from scratch, and it can be stacked up to five times within the same lineage. In the request list, you can check the generation and lineage (source of derivation). Additionally, we have added a score graph to the verification tool "Sensor AD" to help select the decision threshold.

We have released three technical documents for the AI visual inspection platform "DeepAge Sensor" for free. The document titled "How to Launch Visual Inspection AI in a Process with No Defective Samples" (12 pages) summarizes five steps to establish anomaly detection using only good product images, along with pitfalls identified through real data validation. Additionally, we have prepared a product catalog (14 pages) and a "Guide to Integrating AI into Inspection Equipment" (13 pages) for equipment manufacturers. You can download them by providing your name, company name, and email address.

Revamped the DeepAge Sensor's Muon optimizer with Frobenius normalization and BF16 numerical contracts. Validation on 20 real data sets achieved a median maximum accuracy of 100%, with 90% accuracy reached approximately 2.6 times faster than SGD (600 batches vs 1,700 batches). Achieved 97-100% accuracy on high-difficulty data sets that could not be learned with SGD. Operation on the latest NVIDIA GPUs (Blackwell/sm_120) has also been verified.