Skip to content

Research Experience

1. Power plant safety equipment wearing recognition

Duration:06/2020-09/2020

Role:Researcher

Organization:Ningxia SGCC AI and Big Data Innovation Lab

Cooperation:State Grid Corporation of China

  • The research goal of this project is to use deep learning object detection algorithms to identify weather the working wearing safety equipment right or not when entering a construction site.
  • Analysis shows that the recognition requires fast response time and edge calculation, and low mAP is acceptable if the binary judgment is accurate enough.
  • Collected data from cameras and used LabelImg from Github to label the images and resize to 300x300. After data augment, the dataset finally has 4 classes and about 3000 images each class. And transfer them into TFrecords.
  • SSD-300(VGG-based) is Used to detect the safety equipment. Train the model in tensorflow2.1.0 but the loss was not convergent. The reason is in loss function the denominator should be the number of the positive labels and negative labels, but I mistakenly used batch size as the denominator.
  • The model can draw the boundaries of the helmet, reflective vest, insulating gloves and insulating boots. I defined the recognition as valid when the map is greater than 50%. The recall score is 0.88 and the speed is 43fps on my single GPU (Nvidia 2070s) machine.

2. Traffic Targets Detection Based on Deep Learning

Duration:09/2017-01/2018

Role:Postgraduate student

Organization:Institute of Transportation System Science and Engineering (TSE)

Cooperation:Beijing Jiaotong University

  • The research goal of this project is to use deep learning object detection algorithms to identify and track pedestrians, cars, traffic signs and other common targets (13 classes totally) on roads.
  • Collected photos by Python web crawler and camera to expand the VOC2007 and VOC2012 dataset.
  • Enhanced the dataset by rotating, flipping, and adjusting exposure of the photos by a Python lib named Albumentations.
  • Improved the accuracy and speed of feature extraction net, which is GoogLeNet, by adding Space Pyramid Pooling, Depthwise Separable Convolutions and Batch Normalization, and trained it with the enhanced database.
  • Took the improved GoogleNet as the basenet for Faster RCNN and YOLO V2.
  • Adjusted parameters such as Anchor Size, and hyper-parameters such as Learning Rate, Batch Size through experimental comparison to obtain result optimization.
  • Using the captured video to verify on my experimental server. Compared with the original Faster RCNN, the optimized network has improved mAP by 1.9%, increased processing speed by 8 fps. Compared with the original YOLO V2, the optimized network is in mAP Increased by 3.1%, and increased by 4 fps in processing speed.

3. Nuclei Segmentation in Digital Pathology Using Deep Machine Learning

Duration:06/2017-09/2017

Role:Postgraduate student

Organization:Department of Computer Science

Cooperation:University of Birmingham

  • Exploited deep learning to identify breast cancer nuclei.
  • Used Matlab to split the original image dataset, generated 1.2 million 32x32 patches with labels (0-the most central pixel is non-nuclei, 1-the most central pixel is nuclei), and grouped them by 5-fold cross validation.
  • Implemented Alexnet and VGGNet in Tensorflow and classified each pixel with sliding window and small patches.
  • Achieved a pixel level nuclear identification rate (recall rate) of 88%+ through VGG and Alexnet models.
  • Proposed improved sensory field (1x1 improved to 2x2), and the performance of pixel level classification is improved without the increase of calculation cost, the identification rate increased 4%.

4. Brain MRI Data Diagnosis Based on K-means algorithm

Duration:06/2017-09/2017

Role:Postgraduate student

Organization:Department of Computer Science

Cooperation:University of Birmingham

  • Exploited deep learning to identify breast cancer nuclei.
  • Used Matlab to split the original image dataset, generated 1.2 million 32x32 patches with labels (0-the most central pixel is non-nuclei, 1-the most central pixel is nuclei), and grouped them by 5-fold cross validation.
  • Implemented Alexnet and VGGNet in Tensorflow and classified each pixel with sliding window and small patches.
  • Achieved a pixel level nuclear identification rate (recall rate) of 88%+ through VGG and Alexnet models.
  • Proposed improved sensory field (1x1 improved to 2x2), and the performance of pixel level classification is improved without the increase of calculation cost, the identification rate increased 4%.

5. The Prediction on Electric Vehicle Battery Mileage

Duration:02/2016-07/2016

Role:Postgraduate student

Organization:Institute of Transportation System Science and Engineering (TSE)

Cooperation:Beijing Jiaotong University

  • Counted and pre-processed the data of charge and discharge of the electric vehicle battery under different accelerations, temperatures, and operational impairments through the observation and recording by the sensors in the laboratory.
  • Built linear regression to model and predict the power changes of the electric vehicle battery.
  • Conducted real-world test and compared the results with those from the laboratory experiments; found a large gap between the results due to four impact factors.