Deep Learning-based Image Classification of Faulty Ethernet Cables

Objectives

1. To train the Squeezenet model on the Ethernet cables images to classify the status of Ethernet cables on the device under test
2. To create the hardware-optimized inference engine with the trained model using the Intel OpenVINO toolkit for the Intel CPU and Intel-Movidius VPU
3. To develop the Python/OpenCV-based application to classify the real-time status of Ethernet cables in the webcam images

Hardware Platform

Training Node

Hardware

CPU: Intel Core i7-8700K CPU @ 3.70GHz × 12
GPU: GeForce GTX 1080 x 2
RAM: 32 GB
OS: Ubuntu 16.04.5 LTS
Refer to this link for the training platform setup

Software

Intel-Optimized Caffe
Intel OpenVINO Toolkit

Edge Node

Hardware

CPU: Intel Xeon CPU
RAM: 16 GB

Software

Intel OpenVINO Toolkit

Dataset

525 images per class

Output Classes

ON: All the Ethernet cables are ON
OFF: At least one Ethernet cable is OFF

Sample Images

ON

OFF

Preparing the training and validation datasets

1. Split the images into training and validation datasets and create the corresponding text files using createTextFiles.py
2. Resize the images into 227x227 and create the LMDB files using convert_txt_to_lmdb.sh
3. Input the location of training and validation LMDB files in the train_val.prototxt to train the model

Training

Train the fine-tuned Squeezenet model on the dataset using two GPUs

$CAFFE_ROOT/build/tools/caffe train --solver solver.prototxt --weights squeezenet_v1.0.caffemodel --gpu all

Execution Time: 1h 3m 40s; Iterations: 50000

Training and Validation Performance

Building Optimized Inference Engine on Edge Node

Building the optimized inference engine for the Intel CPU

python3 /opt/intel/openvino/deployment_tools/model_optimizer/mo.py --framework caffe --data_type FP32 --input_shape [1,3,227,227] --input data --mean_values data[104.0,117.0,123.0] --output prob --input_model train_cp_iter_50000.caffemodel --input_proto deploy.prototxt --output_dir ./

Building the optimized inference engine for the Intel VPU

python3 /opt/intel/openvino/deployment_tools/model_optimizer/mo.py --framework caffe --data_type FP16 --input_shape [1,3,227,227] --input data --mean_values data[104.0,117.0,123.0] --output prob --input_model train_cp_iter_50000.caffemodel --input_proto ./FP16/deploy.prototxt --output_dir ./FP16/

Testing on Edge Node

Image Classification on CPU

/home/puzzle/Documents/CheckPoint/Classification/ImageClassification/image_classification_sync -d CPU -m /home/puzzle/Documents/CheckPoint/Classification/FP32/train_cp_iter_50000.xml -i /home/puzzle/Documents/CheckPoint/Classification/test_input.jpg

Inference Time: ~3 ms

Image Classification on VPU

/home/puzzle/Documents/CheckPoint/Classification/ImageClassification/image_classification_sync -d MYRIAD -m /home/puzzle/Documents/CheckPoint/Classification/FP16/train_cp_iter_50000.xml -i /home/puzzle/Documents/CheckPoint/Classification/test_input.jpg

Inference Time: ~15 ms

Real-time Image Classification on the webcam video on CPU

/home/puzzle/Documents/CheckPoint/Classification/VideoClassification/video_classification_async -d CPU -m /home/puzzle/Documents/CheckPoint/Classification/FP32/train_cp_iter_50000.xml -i /dev/video0

Real-time Image Classification on the webcam video on VPU

/home/puzzle/Documents/CheckPoint/Classification/VideoClassification/video_classification_async -d MYRIAD -m /home/puzzle/Documents/CheckPoint/Classification/FP16/train_cp_iter_50000.xml -i /dev/video0