Enabling Real-Time Face Expression Classification using Intel® OpenVINO™

Published:06/27/2019   Last Updated:06/27/2019


Face recognition has been used in a broad range of applications such as Security Systems, Marketing and Social Media, for a long time. With the increase of model complexity and hardware technologies a new era of face recognition has begun: Facial Expression Recognition. Deep learning has become essential for achieving state-of-art levels of accuracy and providing robust solutions for recognizing expressions even in different conditions of brightness, contrast and image quality.

facial expressions classifier

This paper focus on the inference optimization process of a facial expression recognition system based on InceptionV3 and MobileNet architectures. It uses Intel® OpenVINO™ to enable real-time applications perform classifications using Deep-Learning models. Two experiments are defined:

  1. Inference using InceptionV3 architecture in Intel® Core™ i7 and Intel® Xeon® 8153.
  2. Inference using MobileNet architecture in Intel® Core™ i7 and Intel® Xeon® 8153.

Solution Architecture and Design

The solution is aimed at classifying the face expression class. The block diagram is shown below:

block diagram


OpenVINO™ is a toolkit that allow developers to deploy pre-trained deep learning models. It has two principal modules: A Model Optimizer and the Inference Engine. Check Install Intel® Distribution of OpenVINO™ toolkit for Linux*1 for more information on how to install the SDK.

Model Optimizer

A set of command line tools that allows you to import trained models from many deep learning frameworks such as Caffe*, TensorFlow* and others (supports over 100 public models)

  • Transform the model into an intermediate representation (IR) to allow the usage of the Inference Engine.
  • Model conversion: Fuse operations, apply quantization to reduce data length and prepare the data with channel reordering.

Inference Engine

Uses an API based code to do inferences on the platform of your choice: CPU, GPU, VPU, or FPGA.

  • Execute different layers on different devices
  • Optimize execution (computational graph analysis, scheduling, and model compression)

Steps to Enable OpenVINO™ Using a TensorFlow* Model

  1. Convert the model to an Intermediate Representation (IR)
  2. Pre-process the image
  3. Setup the Inference Engine code to run the IR.

Creating OpenVINO™ representation

Step 1. Convert the model to an Intermediate Representation (IR)

 /opt/intel/computer_vision_sdk/deployment_tools/model_optimizer/mo_tf.py --input_model frozen.pb \
--input_shape [1,299,299,3] \
--data_type FP32
# The following files will be created:
# frozen.bin
# frozen.xml

def pre_process_image(imagePath):
# Model input format
n, c, h, w    = [1, 3, 299, 299] (InceptionV3)
image         = Image.open(imagePath)
processedImg  = image.resize((h, w), resample=Image.BILINEAR)    

# Normalize to keep data between 0 – 1
processedImg  = (np.array(processedImg) - 0) / 255.0

# Change data layout from HWC to CHW
processedImg  = processedImg.transpose((2, 0, 1))
processingImg = processingImg.reshape((n, c, h, w))
return image, processingImg, imagePath

Step 2. Pre-process the image

# Plugin initialization for specified device and load extensions library if specified. 
# Devices: GPU (intel), CPU, MYRIAD
plugin = IEPlugin("GPU", plugin_dirs=plugin_dir)

# Read IR
net = IENetwork.from_ir(model=model_xml, weights=model_bin)
assert len(net.inputs.keys()) == 1,
assert len(net.outputs) == 1, 
input_blob = next(iter(net.inputs))
out_blob = next(iter(net.outputs))
# Load network to the plugin
exec_net = plugin.load(network=net)
del net
# Run inference
image, processedImg, imagePath = pre_process_image(fileName)
res = exec_net.infer(inputs={input_blob: processedImg})

# Access the results and get the index of the highest confidence score
res = res['dense_2/Sigmoid']
idx = np.argsort(res[0])[-1]

Hardware Configuration

The following are the hardware configurations used for the experiments:

Intel® Xeon® Platinum 8153 processor Intel® NUC7i7BNH
Architecture: x86_64 Architecture: x86_64
CPU op-mode(s): 32-bit, 64-bit CPU op-mode(s): 32-bit, 64-bit
Byte Order: Little Endian Byte Order: Little Endian
CPU(s): 64 CPU(s): 4
On-line CPU(s) list: 0-63 On-line CPU(s) list: 0-3
Thread(s) per core: 2 Thread(s) per core: 2
Core(s) per socket: 16 Core(s) per socket: 2
Socket(s): 2 Socket(s): 1
NUMA node(s): 2 NUMA node(s): 1
Vendor ID: GenuineIntel Vendor ID: GenuineIntel
CPU family: 6 CPU family: 6
Model: 85 Model: 142
Model name: Intel® Xeon® Platinum CPU 8153 @ 2.00GHz Model name: Intel® Core™ i7-7567U @ 3.50GHz


4 Stepping: 9
CPU MHz: 1800 CPU MHz: 4000
BogoMIPS: 4000 BogoMIPS: 7000
L1d cache: 32K L1d cache: 32K
L1i cache: 32K L1i cache: 32K
L2 cache: 1024K L2 cache: 256K
L3 cache: 22528K L3 cache: 4096K

Software Used

The following is the software configuration used:

OS CentOS* Linux release 7.4.1708 (Core)
Kernel Version kernel 3.10.0-693.el7.x86_64
Python* Version Python* 3.6.1
TensorFlow* Version 1.10
Anaconda* Version 4.3.25
OpenVINO™ SDK Version 2018.3.343


The first assessment done within OpenVINO™ Toolkit was based on InceptionV3 topology. The results demonstrated an increase up to 7.12x improvement in inference time running on integrated Intel Graphic Processing Unit (iGPU) of an Intel® NUC7i7BNH.

results demonstrated an increase up to 7.12x improvement

For the MobileNet topology, the inference process had a performance improvement by 18.33x, as the topology is lighter than InceptionV3.

for MobileNet the performance improved by 18.33x

To take advantage of the full resources of the CPU, multi-inference was used to share the workload across the cores to reduce memory overhead and thereby increasing throughput and decreasing inference time by effective parallelization. After further optimization with OpenVINO™ Toolkit, the performance improved to 25.85X using multi-inferences (up-to 16 processes at the same time on a single node).

# Execute multi-inference according to the number of cores available in the hardware (in this case, 64 cores are available)
CMD="python yourScript.py” 
# 4 Cores / process(16)
numactl -C 0-1,2-3     $CMD & numactl -C 4-5,6-7     $CMD &
numactl -C 8-9,10-11   $CMD & numactl -C 12-13,14-15 $CMD &
numactl -C 16-17,18-19 $CMD & numactl -C 20-21,22-23 $CMD &
numactl -C 24-25,26-27 $CMD & numactl -C 28-29,30-31 $CMD &
numactl -C 32-33,34-35 $CMD & numactl -C 36-37,38-39 $CMD &
numactl -C 40-41,42-43 $CMD & numactl -C 44-45,46-47 $CMD &
numactl -C 48-49,50-51 $CMD & numactl -C 52-53,54-55 $CMD &
numactl -C 56-57,58-59 $CMD & numactl -C 60-61,62-63 $CMD &

performance improvement by 25.85x

The same optimization was repeated using the MobileNet topology where the gains were even better, reaching a speed-up of 95.64X using multi-inferences.

a speed-up of 95.64X using multi-inferences


This paper showed that is possible to speed-up the inference process by 95.64x using OpenVINO™ toolkit and Intel® Xeon® Platinum 8153 processor. It covered the steps required to transform the original model to an optimized model and provided a sample of the pre-processing and inference script. Moreover, the usage of multi-inference increases the throughput resulting on inference time reduction.


1. Install Intel® Distribution of OpenVINO™ toolkit for Linux*

2. Model Optimizer Developer Guide

Product and Performance Information


Performance varies by use, configuration and other factors. Learn more at www.Intel.com/PerformanceIndex.