OCR system for financial sheets [Oct. 2020 – Dec. 2020]
The project built an OCR system for financial sheets analysis. My main contribution: 1. Train and deploy a text recognition model based on CRNN; 2. Adopt an data augmentation strategy inspired by the paper EDA to handle the small sample size problem. The avg. time for text detection and recognition for an image (1000x2000px) is within 5s on CPU and the recognition accuracy over ~10,500 test samples is above 95%.
ID-card rectification [Jun. 2020 – Jul. 2020]
The project realized an image rectification API for ID-card images in natural scenes. It mainly utilized an edge detection network for card contour detection and opencv interface for perspective transformation. The API provide single and batch image processing services and currently serves for RightPrint Mini Program.
Fabric defect detection [Jun. 2019 – Aug. 2019]
An algorithm developed for fabric defect detection. The algorithm goes through the following step: 1. Image preprocessing (crop, resize, remove uneven illumination, and remove moire patterns); 2. Surface defect detection (OTSU threshold, morphological transformation); 3. Linear defect detection (Canny edge detection, Hough line detection); 4. Defect feature extraction and classification; 5. Display and generate for XML output.
Predicting radiotherapy sensitivity of laryngeal cancer based on DNNs [Jul. 2019 – Oct. 2019]
In this study, we enrolled 200 patients with laryngeal cancer (LC) who underwent standard radiotherapy alone. Patients were followed up and were classified into radiotherapy-sensitive (LC-RS) and radiotherapy-tolerant (LC-RT) groups according to their prognosis. I took the responsibility of modeling a convolutional neural network based on GoogLeNet, VGG16 and ResNet50 to predict the sensitivity of patients with radiotherapy, and combined the clinical features such as EBV, tumor markers, etc. to compare the imaging difference between LC-RS and LC-RT. Experimental results showed that our model reached 74.5% prediction accuracy among 55 patients of CT scans.
Medical MRI Segmentation [Sep. 2018 – Jun. 2020]
Work#1 (ICIP-19): A LSTM method with Multi-modality and Adjacency constraint is proposed for segmenting three tissues in Brain MRI. Two feature sequence generation ways in the method are used, i.e., features with pixel-wise and superpixel-wise adjacency constraint. The method is more robust than clustering-based methods like FCM, K-Means, and performs better than other feature classifiers like SVM and KNN, achieving 98.66% Dice Coefficient on the BrainWeb dataset.
Work#2 (AAAI-20): It proposes a Recurrent Decoding Cell (RDC) for hierarchical feature fusion in the encoder-decoder segmentation networks. RDC leverages the ability of convolutional RNNs in memorizing long-term context information. The RDC-based segmentation network achieves 99.34% Dice Coefficient on the BrainWeb dataset, which is better than FCN, SegNet and U-Net, and is robust to image noise and intensity non-uniformity in medical MRI.
My spirit (Unity Games) [Dec. 2015 – May. 2016]
My Spirit is a 2D hand-drawn-style adventure game. It tells the story about the legendary life of Jerry from the spirit world to the real world. Players use keyboard or Xbox controller to control the characters. We use Unity as the game engine and C# as the programming language. The game is produced by our team EBM, I took charge of the character/UI design, shooting and part of the algorithm realization in this project.
FarmKit [Dec. 2014 – May. 2015]
FarmKit is a software and hardware framework made specially for agricultural research. It consists of three parts: Intelligent Agricultural Control Terminal, Robot Surveillance and Control System and the Greenhouse Control System. I tested the performance of the robot in greenhouse and collected environment data. This project competed in Imagine Cup 2015 World Citizenship Group and won the second prize in national final.