An artificial intelligence solution for atypical defect detection that meets the needs of decision-makers seeking to optimize production speed and reduce costs.
Class Equalizer Support
When the data ratio between classes is imbalanced due to a lack of edge cases, the Class Equalization function automatically adjusts the contribution of each class during training. This mitigates training bias toward specific classes and effectively improves overall inference stability.
Fast Processing Speed
The FLImaging® AI framework demonstrates higher performance than PyTorch and TensorFlow, the deep learning frameworks primarily used by existing AI developers. Based on this, it greatly assists AI research and development by producing results faster. Furthermore, its high performance provides the relative effect of enjoying premium GPU capabilities at a lower GPU cost.
Easy Label Data Management
The .flif extension, dedicated to FLImaging®, provides a structure capable of integrating and storing both image data and labeling information (Image Figure). Additionally, it supports Multi-Page functionality, allowing multiple training images to be managed collectively in a single file, thereby significantly enhancing data consistency and management convenience.
One-Click Pre-trained Application
The FLImaging® AI framework supports a flexible pre-trained learning structure. Data pre-trained in FLNet®, a high-performance classification model, can be loaded into FLSegNet®, an FLNet® Backbone-based Semantic Segmentation model, ensuring fast convergence and high accuracy even with a small amount of data. This allows you to simultaneously achieve shorter training times and improved performance.
The .flif extension, dedicated to FLImaging®, provides a structure capable of integrating and storing both image data and labeling information (Image Figure). Additionally, it supports Multi-Page functionality, allowing multiple training images to be managed collectively in a single file, thereby significantly enhancing data consistency and management convenience.
Real-time Object Augmentation Exclusive to FLImaging® AI
FLImaging® AI provides various types of augmented training images based on its rich augmentation features. Furthermore, its Object Augmentation function allows for precise augmentation at the object level rather than the entire image, effectively expanding data diversity and improving training efficiency. Based on this, a high-performance model can be trained even with a limited amount of training data. All image augmentation results are dynamically reflected in the training data in real-time during the training process without significantly affecting the training speed.
Various Object Augmentation Examples
(1) Original
(2) Original -> Simple Quantity Augmentation
(3) Original -> Quantity + Size + Pose Augmentation
(4) Original -> All Augmentation Options Used
FLImaging® Development Studio
Current progress can be tracked in real-time via the training progress information window.
Training can be effortlessly performed just by clicking the Learn button in the GUI framework after setting the Learn Image, Validation Image, and Epochs. In addition to the default settings, users can finely tune various parameters such as Learning rate, Weight decay, Mini-batch size, and Augmentation. It supports an Auto Mini-Batch mode; instead of users having to manually find the optimized value, the FLImaging® AI framework automatically sets the optimal Mini-Batch size according to the number of images to be trained, leading to faster training speeds and more precise results.
Real-time Training Information Monitoring
Through the real-time monitoring function, past and present state changes during training can be easily checked via the GUI framework.
Various High-Performance Options and User Convenience
Original Resolution Training Support
Provides the ability to train at the original resolution, enabling the detection of extremely small defects down to 2~3 pixels in size.
Parallel Processing Support
Delivers optimized performance tailored to the user's environment by supporting SIMD, Multi-threading, and CUDA.
Multi-GPU Support
Simultaneous training and inference can be executed via up to 64 GPUs to train massive amounts of images at ultra-high speeds. Developers do not need to write complex multi-GPU programming codes to utilize multi-GPU functionality.
“FLImaging® Guarantees Success.”
The fastest deep learning, made easiest.
EXCELLENCE
Our passion and dedication to excellence translate directly into productivity at our customers' sites.
Save and Load Training Information at the Optimal Point
It is possible to save and load training information. It can be automatically saved under the best validation or cost conditions, ensuring you obtain the optimal model before overfitting occurs.
Configurable Termination Conditions
Supports an automatic termination function when training reaches a predefined state, such as a set number of epochs or a target cost.
Various Optimizers Provided
Supports users in achieving their desired training outcomes by conveniently allowing the configuration of various optimizers and related parameters.
Proprietary AI Framework
Without the need to install Nvidia's cuDNN library, you can immediately utilize the high-performance deep learning solution developed directly by Fourth Logic with just a single click of the FLImaging® library setup button.
>
Product
>
FLImaging®
>
