DisplacementMLP+: Unsupervised Neural Network for Dynamic Scene Analysis
Daria Mangileva, Никита Сергеевич Марков · 2025
The multilayer perceptron (MLP) represents one of the earliest models in neural network architecture. In recent years, there has been a resurgence of interest in MLPs, driven by the emergence of innovative technologies that address the challenges associated with underfitting in these models. In particular, there has been a rise in studies that involve the analysis of dynamic scenes using multilayer perceptrons (MLPs). Considered model architecture includes two MLPs: one serves as an image generator Ï, while the other functions as a grid generator $\ddot G$.Our previous research has shown that this model configuration, with optimally chosen parameters, outperforms state-of-the-art methods on images characterized by gradient spatial variation in pixel values. Based on these results, this paper introduces the DisplacementMLP+ neural network model, which uses a novel approach to train an image generator. The quality of this model was evaluated using a sample of 485 artificially distorted medical and biological images. DisplacementMLP+ showed a 16% improvement in the quality of the displacement field calculation compared to the traditional MLP-based model used to analyze dynamic scenes. The performance of the proposed model is comparable to the state-of-the-art neural network DICNet-corr. In addition, DisplacementMLP+ showed advantages over DICNet-corr in terms of maximum spatial mean square error, especially for images representing open, non-isolated hearts. Dynamic analysis of such images is crucial for further visualization of the mechanical spiral waves that occur during arrhythmia, which can improve our understanding of arrhythmogenic processes.