• Mastered algorithmic logic and interfacing between software and ML algorithms, gaining in-depth insights into the entire system flow from image capturing to output. This expertise facilitated the creation of detailed system diagrams, improving cross-departmental understanding and project alignment.• Identified and established essential requirements for simulation tools and parameters, enabling effective analysis while adhering to GDPR (General Data Protection Regulation) and strict IP (Intellectual Property) regulations. Demonstrated capability in devising evaluation tools to supplement the analysis process.• Spearheaded performance analysis and optimization for four high-profile Xiaomi projects, leading to successful market launches. Identifying and addressing corner cases and unexpected performance issues through deep algorithm understanding and image scenario familiarity. • Expertly managed long-term data analysis under diverse scenarios for performance comparison and validation during version releases. Proactively tracked performance distributions, making strategic recommendations for enhancements, and accurately deducing user behavior to pinpoint root causes in complex scenarios.

Optical character recognition (OCR) algorithm development on an algorithm model generation platform product • Provided algorithm scheme and ONNX models for various projects including label text, wafer character, chip lettering, shank engraving, and cigarette box engraving recognition. Proposed and completed feasible improvements inspirated during projects such as flexible support for character range and filter box, automatic parameter setting based on training set, visualized display of false alarm on a separate report page as guideline to iterate the algorithm built by the platform.• Collected relevant datasets and font libraries to synthesize half a million industrial images, creating a training set with strong generalization ability to improve effect of universal character recognition in industrial scenarios so that realizes the function of roughly recognizing characters without labeling and few shot training for most industrial recognition tasks.• Maintained Character recognition algorithm module and implement product-side function such as transport protocols with backend, performance test establishment, secure docker image authentication, conformance check and performance correlation for optical character recognition algorithm on the platform.• Developed a small OCR algorithm platform using Pyinstaller and Pyqt for internal use which can be quickly installed on local computers with packaged python environment and visualized operation pages. This small software is equipped with functions including synthetic training image set, image annotation, optical character recognition model training and testing, which can also flexibly increase module performance with extra data augmentation buttons and postprocessing buttons.

Shenzhen Shennan circuit PCB board defects-detection algorithm scheme delivery• Adjusted the algorithm scheme according to the change of requirements and on-site feedback; be responsible for evaluation set establishment, effect analysis, data cleaning, pre-processing and post-processing adjustment, model optimization and performance comparison. • Communicated with the back end of the platform to realize the function of joint debugging and protocol transmission; established a conformance test process, proposed, and implemented the algorithm encryption function. • Liaised with the project leader from Shenzhen Shennan circuit PCB; managed requirements gathering. Coordinated support personnel and promoted the process and documentation of reporting for daily progress. In the pre-launch security testing stage, lead team to respond on-site, responsible for effect analysis, task splitting, data preparation, model training, release reviewing, and effect interpretation.

• Conjugated a custom-built detection network with a classification network to develop an Optical Character Recognition (OCR) algorithm for images taken by a product under development.• Enhanced the detection network by incorporating a feedback layer, conducting experiments to analyze the impact of feedback positioning, initialization modes, and other variables on model performance.• Redesigned the preprocessing and post-processing components of the algorithm, eliminating reliance on TensorFlow to better tailor the custom model.• Defined optimal training set sampling criteria and segmented detection results into positioning boxes, aligning outputs with the classification algorithm to fulfill advanced application requirements.

• Researched algorithms for pixel calibration and precision measurement of binocular vision.• Visualized interpolation functions of Bilinear Interpolation and Bicubic Interpolation, and simulated subpixel correlation coefficient function model by projecting Band-limited Random.