<p>This paper introduces a lightweight and precise convolutional neural network (CNN) model designed for mobile edge computing devices. This model operates primarily in the spectral domain, making it highly efficient for artificial intelligence tasks. Unlike conventional models, it performs a single domain transformation on the source feature map, while the remainder of the network processes data exclusively in the spectral domain. To further enhance performance, specialized spectral size and depth optimization techniques have been applied, significantly reducing computational complexity. The proposed model was rigorously tested on a newly developed 94-class ASCII character dataset, which includes a diverse range of lowercase and uppercase letters, numbers, symbols, and special characters across various fonts. Detailed comparisons were made between the proposed spectral CNN models and equivalent models operating in the conventional spatial domain. The results demonstrate that the accuracy of the spectral and spatial models is comparable. However, the spectral model outperforms in other critical metrics, such as precision, where it shows an improvement of over 95%. Moreover, the computational workload in the spectral domain is reduced by several orders of magnitude. For instance, similar models, such as VGG7, exhibit a 2.1x faster training time and a 4.4x faster testing time when implemented in the spectral domain. Compared to existing methods in the literature, particularly for the MNIST dataset, the proposed model not only achieves the highest accuracy at 97.5% but also demonstrates up to a 20x reduction in computational workload. </p>