The Neural Network Simulation of Instance-Aware Semantic Image Segmentation System

This paper presents the system of the recognition and segmentation of objects which is based on an effective method for semi-controlled instance-aware segmentation of video and images. The system allows to achieve high quality forecasting of segmentation annotations for video frames while processing in real time. For this purpose a neural network model was proposed in the form of a Siamese deep network with a structure “encoder-decoder” which is designed to take advantage of both propagation-based and detection-based methods while avoiding the weaknesses of both approaches. The network was trained in a two-stage training process that uses both synthetic and real data and works reliably without online training or post-processing. The quality of the trained model was tested on training video.