• Design of deep neural network for large scale face recognitionImprovement of the performance by combining arcface loss and contrastive loss.Design of a masked face recognition network using attention mechanism.• Design of face clustering and search algorithm using multi-core CPU and multi-GPUImplementation of a face clustering algorithm using similarity search method by Hierarchical Navigable World graph (HNSW).• Construction of People Surveillance System using human tracking and face recognitionIntroducing of deep learning based human tracking module like ByteTrack.Implementation of accurate face detection module by ensemble of MTCNN and RetinaFace based Resnet50 backbone.Construction of a vector database for storing face & human features.Implementation of face & human search algorithm using ANN search methods.• Implementation of super-fast audio filter algorithm in SIMD & CUDA C++Implemented 1D convolution operation of super-long audio signal vector in CDUA C++ and achieved ~10x speed improvement on GTX 1060 GPU compared to AVX SIMD implementation on Intel Core i5 CPU.• Implementation of license plate recognition algorithm in CUDA C++Implemented a resampling-based cascaded framework decoupling a license plate detection model and a license plate recognition model using weight-sharing classifier.• Design and deployment of models for image/video segmentation on mobile devices in real-time.Designed lightweight image semantic segmentation network using residual dense blocks and dilated convolutional pyramid components and then boosted the performance by using boundary information and semantic connectivity information.For lightweight video segmentation, designed an expanded segmentation network using previous predicted segmentation information.Besides, boosted the accuracy of student (final) model by introducing channel-wise knowledge distillation.• Design of human pose estimation networkDesigned a lightweight 2D pose estimation network using lightweight bottleneck block and channel-only & spatial-only self-attention mechanism.• Design of image super resolution networkDesigned an image super-resolution network using progressive multi-scale residual dense block, especially channel & pixel-wise attention mechanism for finding the inherent correlations among image features.