Analysis of Object Phase Transition Process Based on Image Segmentation and Graph Neural Network

Based on Canny operator edge detection and watershed algorithm, this paper establishes the image segmentation model. Then, based on the open source OCR (Optical Character Recognition) model, FPN-PSENet is used to reach the regional detection on the text, and CRNN plus CTC method is used to screen as well as intercept the identified text. According to the obtained data, the function diagram of temperature and time is achieved, and the reasons for the inaccurate data of some curves are summarized after comparative analysis. Additionally, by introducing colour eigengradients and position hue bias functions, and defines the structural similarity indices. When further modelling the temporal features, the DETR (Detection Transformer) neural network is improved to be used in four ways. A prior decision is used to classify the characteristics of different stages: the luminance, the contrast, the contour, the centre-edge difference and other explicit/implicit features, decoupling the excessive coupling due to different importance of different features. The results indicates that the scattering plots converge to the practical curve, which shows the strength of the model.