Abstract:
Face recognition systems have many useful applications in real life scenario, such as authentication for personal devices, border control at airports or attendance control for classrooms or institution. For these use cases, it is important for the system to achieve high recognition performance while still be able to process a large amount of transactions. To ensure that low quality images do not negatively contribute to the biometric performance, a method to quantify the quality of images taken from biometric samples is highly desirable. The current draft international standard of ISO/IEC 29794-5 introduces the concept of component quality, which quantitatively expresses the quality of a given biometric sample. In this work, we look into NeutrEx – a recently proposed quality measure which quantifies the expression neutrality of facial images in the context of ISO/IEC 29794-5. Additionally, we optimize the NeutrEx model to achieve better efficiency with regard to number of parameters, storage space, and inference time. Our proposed model NeutrEx-lite is intended to be a more streamlined version of NeutrEx, such that it becomes more suitable for real life applications, while still maintain respectable performance relative to the original one. In order to achieve this, we research well-known methods in the field of neural network optimization; such as pruning, quantization and knowledge distillation and apply them to NeutrEx. Our contributions are as follow:
- Analysis of NeutrEx with regard to the underlying architecture to identify
possible optimization spots. - Research into possible optimization techniques in the literature that can help us increase the efficiency of NeutrEx.
- Concrete implementations of such techniques and benchmark the results against the original NeutrEx.
- Observations and discussion of final result and possible future works with regard to NeutrEx and NeutrEx-lite.