Pass/Fail, grade quality, locate defects, etc.
Locate people, measure cell density, detect obstacles, etc
Airplane, cars, people, cells under a microscope, etc.
Semi conductor inspection, QC of printed materials, etc
- Novelty detection
- Anomaly detection
- Image contextual segmentation
- Tracking with reinforced learning as the target changes
- Stereoscopic distance evaluation
- Edge detection
- Noise removal
- Image compression
- From annotations made by domain experts on collection of reference images
- From live interrupts triggered by an operator
Non supervision training
- To re-inforce a knowledge getting less confident
- To synchronize knowledge between stereoscopic engines
- To build a dictionnary of viewlets describing reference images for compression or segmentation
- To build a reactive image memory frame
Train multiple networks to make more robust decision
- Experts can be trained to be redundant to minimize false positive
- Experts can be trained to be complementary to waive uncertainties wisely
Multiplicity of outputs
- List of identified objects (location, recognized categories, confidence levels)
- Transform images (maps of categories and confidence)
Read more about the CogniSight SDKs
CogniSight engines can be instantiated in MPU and/or FPGA and access the NeuroMem neurons through a variety of communication bus (SPI, Quad SPI, I2C, etc) or directly through the NeuroMem parallel bus which delivers optimum recognition speed (>50,000 recognitions per second for vectors of length 256).
The NeuroShield HDK for Xilinx ZYNQ SoC is the perfect tool to interface to the NeuroMem neurons through an ARM processor and/or an FPGA.