Low- and High-Frequency Analysis of Flower Images Using FFT and DCT

Abstract

This paper presents a comprehensive frequency-domain analysis of a flower image dataset comprising 733 images across 10 botanical classes. We investigate the discriminative capacity of low- and high-frequency spectral features extracted via 2D Fast Fourier Transform (FFT) and 2D Discrete Cosine Transform (DCT). A complete preprocessing pipeline is established, including ImageNet-normalized resizing to 224×224 pixels, stratified 70/15/15 train-validation-test splitting, and augmentation strategies suitable for deep learning. Frequency band energy is computed for low (DC to 10%), mid (10–35%), and high (above 35%) spatial frequency regions. Results consistently show low-frequency band dominance across all flower classes in both FFT (92–96% energy) and DCT (94–98% energy) decompositions. Class discriminability is assessed via silhouette scores following PCA reduction; the DCT low-frequency 32×32 block yields the best result (−0.1254) compared to FFT radial power (−0.1716) and the full DCT 64×64 block (−0.2485). The negative scores indicate significant inter-class visual similarity in frequency space, motivating the use of learned CNN features rather than handcrafted frequency descriptors for fine-grained flower classification. All code, preprocessing pipelines, and visualizations are documented and reproducible.