L2-Constrained Softmax Loss for Face Verification
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Face verification systems using deep convolutional neural networks (DCNNs) face performance gaps due to softmax loss not optimizing features for higher similarity scores for positive pairs and lower scores for negative pairs, particularly in unconstrained settings with extreme variations in viewpoints, resolution, occlusion, and image quality, caused by data quality imbalance and bias towards high-quality faces.
Innovation Solution
Implementing an L2-normalizing layer to constrain feature descriptors to lie on a hypersphere of a fixed radius, which enforces L2-softmax loss, ensuring all features have the same norm, thereby providing equal attention to both good and bad quality faces and strengthening the verification signal by maximizing the margin between positive and negative pairs.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Ease of manufacture
If softmax loss is used for training face verification system, then it can be easily implemented using inbuilt functions and converges quickly, but it is biased to the sample distribution and ignores rare difficult faces from training mini-batch
Solution Approach 1:
The patent changes the loss function parameter from standard softmax loss to L2-constrained softmax loss. This parameter change modifies the optimization objective to explicitly maximize similarity scores for positive pairs while minimizing scores for negative pairs, thereby addressing the bias toward high-quality faces while maintaining ease of implementation through existing deep learning frameworks.
2Quantity of substance
If softmax loss maximizes conditional probability of all samples in mini-batch, then it fits well to high quality faces, but it ignores rare difficult faces and does not optimize verification requirement
Solution Approach 1:
The patent applies local quality by differentiating the treatment of different sample types within the mini-batch. The L2-constrained softmax loss function specifically optimizes for positive pairs (same identity) to have higher similarity scores while negative pairs (different identity) have lower scores. This local differentiation ensures that difficult faces are not ignored despite their rarity in the overall dataset.
3Measurement precision
If deep convolutional neural network is trained with softmax loss, then it achieves good performance on curated datasets like LFW, but it fails to achieve similar accuracy on faces with extreme variations in viewpoints, resolution, occlusion and image quality
Solution Approach 1:
The patent introduces dynamics by making the loss function adaptive to different face qualities and conditions. The L2-constrained softmax loss dynamically adjusts the optimization focus based on the verification requirement, maximizing similarity for positive pairs regardless of their quality or difficulty level. This dynamic adaptation enables the system to maintain high accuracy on both curated datasets and unconstrained settings with extreme variations.
Data Source
AI summary
Various face discrimination systems may benefit from techniques for providing increased accuracy. For example, certain discriminative face verification systems can benefit from L2-constrained softmax loss. A method can include applying an image of a face as an input to a deep convolutional neural network. The method can also include applying an output of a fully connected layer of the deep convolutional neural network to an L2-normalizing layer. The method can further include determining softmax loss based on an output of the L2-normalizing layer.