Multimodal Image Fusion and Classification of Power Equipment Using Non-Subsampled Contourlet Transform and Adaptive PulseCoupled Neural Network

Abstract

This paper presents a multimodal image fusion and classification method for power equipment based on the Non-Subsampled Contourlet Transform (NSCT) and Adaptive Pulse-Coupled Neural Network (APCNN). The approach begins with image normalization, geometric alignment, and adaptive noise filtering as preprocessing steps. The NSCT is then applied to decompose input images into low- and highfrequency subbands. Low-frequency components are fused using phase congruency weighting to retain energy features, while high-frequency subbands with structural details are selectively fused using APCNN for precise edge and contour extraction. For efficiency, subbands beyond the fifth decomposition level use local energy maximization for fusion. Experiments were conducted on a dataset of 3,000 images of transformers, current transformers, and disconnectors collected by inspection robots. The model achieved maximum recognition accuracies of 99.39% for transformers, 99.57% for current transformers, and 98.74% for disconnectors. The average classification time per image was 2.36 seconds. Compared with APCNN, PCNN, LeNet, AlexNet, and SVM, the proposed NSCT-APCNN model demonstrated superior performance in accuracy, F1-score, and processing speed. This work provides an effective and scalable solution for real-time multimodal image classification in substation inspection scenarios, with potential for extension to fault detection in smart grids.