A NOVEL HYBRID ALGORITHM FOR OBJECT CLASSIFICATION BASED ON THE INTEGRATION OF MULTIFRACTAL FEATURE EXTRACTION AND DEEP NEURAL NETWORKS

This paper presents a novel hybrid algorithm designed for the intelligent analysis and high-precision classification of complex textured images and non-linear systems. While classical Deep Convolutional Neural Networks (CNNs) are effective in learning the global semantic features of images, they often fail to adequately represent the microscopic characteristics of objects in regions with significant scale variations and severe noise. In this study, object descriptors were extracted using a multifractal analysis framework—specifically, the singularity spectrum algorithm—which mathematically captures the self-similarity, lacunarity, and singularity regularities inherent in natural and technical images or signals. In the proposed model, multifractal invariants were introduced as an additional informational vector into the final fully connected layers of a deep learning network (Feature Fusion). It has been demonstrated that the proposed mathematical approach, when applied to the classification of complex objects (such as vascular networks or aerospace porosities), significantly improves accuracy and sensitivity metrics while optimizing computational load compared to traditional monofractal and deep learning models.