FA-CDDL: Contrastive deep dictionary learning with frequency augmentation

Deep Dictionary Learning (DDL) is a powerful paradigm for visual representation learning, yet prevailing DDL frameworks suffer from three core limitations: (i) they optimize sparse codes locally and underutilize global semantic relations among samples, (ii) they allow features from different categories to collapse in the embedding space, yielding redundant and weakly discriminative encodings, and (iii) they largely ignore frequency-domain cues that carry high-frequency details critical for recognition. To resolve these challenges, we propose FA-CDDL, a Frequency-Augmented Contrastive Deep Dictionary Learning framework. First, we introduce a supervised, multi-layer contrastive learning module that operates at every coding layer. By dynamically forming positive and negative pairsfrom label information, it enforces intra-class compactness and inter-class separation, thereby reducing feature redundancy and sharpening discrimination. Second, we design a frequency augmentation mechanism: encoded features are decomposed in the amplitude spectrum, high-frequency components are retained and reconstructed with phase coherence, and the result is concatenated with spatial features to form enhanced positive pairs. This selectively suppresses low-frequency noise while preserving discriminative detail. Third, we couple reconstruction loss with the supervised contrastive objective in a layerwise optimization loop, writing the learned discriminative signal back into dictionary atoms and codes for globally consistent improvements. Extensive experiments across character, face, and scene/image benchmarks (MNIST, EMNIST, AR Face, Extended Yale B, Scene 15, CIFAR-10, Mini-ImageNet and Places365) show that FA-CDDL consistently outperforms classical sparse coding, single-layer dictionary methods, and recent deep DDL baselines, delivering more compact, robust, and discriminative representations-particularly under frequency-domain perturbations.