A real-valued DCT-based spectral CNN architecture for efficient edge deep learning

Spectral Convolutional Neural Networks (SpCNNs) offer a pathway to computational efficiency by performing convolutions in the frequency domain. While FFT-based SpCNNs reduce convolutional complexity, their reliance on complex valued operations and inverse transforms incurs high memory and latency costs, limiting their utility for embedded and edge applications. This paper proposes an enhanced, real-valued Discrete Cosine Transform (DCT)-based SpCNN that eliminates inverse transforms and complex arithmetic by performing all operations, including convolution, activation, and pooling entirely in the DCT domain. A modified real-valued spectral activation function is introduced to enable effective nonlinearity in frequency space. Experimental evaluation on MNIST and a 94-class ASCII benchmark datasets demonstrates that the proposed model achieves up to 20% reduction in computational workload and 19% reduction in memory access compared to a previous spectral model. Additionally, LeNet5 achieves improved accuracy (98.44%), and the architecture exhibits significantly faster inference and higher energy efficiency in both batch and real time settings. These results establish the DCT-based SpCNN as a practical and scalable solution for deployment in resource constrained systems. • Real-valued DCT-based SpCNN with no inverse transforms or complex numbers. • DCT-domain spectral activation enables full-frequency nonlinear operations. • Reduces FLOPs by 20% and memory cost by 19% versus prior spectral models. • Achieves 98.44% accuracy on MNIST and strong results on ASCII-94 dataset. • Evaluated for latency, power, and throughput in batch and real-time modes.