Chinese researchers create new face biometrics training model for expression recognition

Researchers from Jilin Engineering Normal University in China have developed a new, lighter convolutional neural networks (CNNs) model for facial expression recognition.

The findings were published in the Journal of Electronic Imaging, and describe a model (based on Xception) designed to find a balance between training speed, memory usage, and recognition accuracy.

The original version of the Xception model is 71 layers deep and enables the loading of a pre-trained version of the network trained on more than a million images from the ImageNet biometric database. The pre-trained network can then classify images into a thousand object categories.

When compared to pre-existing CNN models, however, the improved Xception model is different as it uses depth-wise separable convolutions, lead researcher Dr. Jia Tian explains.

Convolutions are the core operation performed at each layer of a CNN. The specific type of convolution used in Tian’s research differs from the standard one as it processes different channels (like RGB) of the input image independently. It then combines the results at the end of the process.

Further, the model combines this type of convolution with a technique called “pre-activated residual blocks,” which reportedly leads to greatly reduced computational costs and number of parameters necessary for accurate classification.

“We managed to obtain a model with good generalization ability with as little as 58,000 parameters,” Tian explains.

According to the research paper, the new model was tested against a number of other facial recognition algorithms in a classroom setting.

All models in the experiment were reportedly trained and tested using the “Extended Cohn-Kanade dataset,” which counts over 35,000 labeled images of faces expressing common emotions.

The model developed by Tian’s team reportedly exhibited the highest accuracy (72.4 percent) in recognizing facial expressions with the least number of parameters.

“The model we developed is particularly effective for facial expression recognition when using small sample datasets,” Tian says.

Moving forward, the researcher said the team intends to build on the findings to further improve the system’s accuracy.

“The next step in our research is to further optimize the model’s architecture and achieve an even better classification performance,” Tian concludes.

Expression recognition technologies are rapidly developing, driven by the increased efficiency of face biometric algorithms.

For instance, in April, a coder created a new computer interface using not-so-subtle facial expressions to complete an apparently real job interview code test.

Days later, Apple was served a lawsuit for allegedly infringing on a facial expression recognition patent for messaging systems.

Article Topics

biometrics  |  expression recognition  |  face biometrics  |  machine learning  |  research and development