Facial Attribute Detection Using Local Region Segmentation

The patent divides the facial image into multiple local regions (e.g., eye region, mouth region, cheek region) and processes each region separately using small classification models. This segmentation allows the system to achieve accurate detection while using minimal memory, as each local model requires far less memory than a single large global model.

The patent transitions from a single global classification model to multiple local classification models operating in parallel across different spatial regions. This dimensional change from one comprehensive model to many specialized models enables both high accuracy and low memory consumption, as each local model can be much smaller while collectively covering the entire face.

By dividing the face into local regions and using small classification models for each region, the patent enables parallel processing that significantly increases processing speed. The segmented approach allows multiple regions to be processed simultaneously, achieving real-time performance while maintaining accuracy.

The patent employs lightweight, simple classification models for each local region that are computationally inexpensive and fast to execute. These simplified models trade some individual complexity for speed, enabling real-time processing when aggregated across multiple regions.

The patent creates a universal framework where the same local region segmentation and classification approach can detect multiple facial attributes (smile, gender, age, etc.) simultaneously. By making the system multi-functional through parallel attribute detection on the same local features, it achieves high versatility without proportionally increasing complexity.

The patent combines multiple attribute detection tasks into a unified framework that processes local regions once and extracts multiple attributes from the same features. This merging of detection tasks reduces overall complexity compared to running separate detection systems for each attribute type.

The segmentation into local regions with small classification models enables deployment on resource-constrained embedded devices while maintaining accuracy. Each local model is small enough to fit in limited memory and execute on low-power processors, yet collectively they achieve high detection accuracy.

The patent changes the parameters of the classification models from large, complex global models to small, simple local models with reduced feature dimensions and fewer parameters. This parameter reduction enables deployment on embedded devices while preserving detection accuracy through the collective power of multiple specialized models.