Multimedia codec evaluation has traditionally relied on the Rate–Distortion trade-off, which captures the relationship between bitrate and reconstruction quality but ignores the operational energy cost of encoding and decoding. This dimension becomes critical when comparing classical and neural codecs, as the latter often reach aggressive bitrate regimes at substantially higher computational cost. This work introduces a Rate–Distortion–Energy (R-D-E) evaluation framework that treats operational energy as a third explicit objective. The benchmark is applied across three domains: image coding (nine codecs or families across Kodak, Tecnick, DIV2K, and CLIC2020), audio coding (Opus, EnCodec, DAC, SNAC, WavTokenizer over speech, ambient sound, and music), and video coding (x264, x265, SVT-AV1, VVenC, and four DCVC variants on UVG in LDP profile). Energy measurements follow a batched protocol that combines CPU (RAPL) and GPU (NVML/Zeus) telemetry, isolates the codec pipeline from perceptual metric computation, and subtracts system idle power. Results show that incorporating energy substantially reshapes codec rankings, with energy dynamics spanning over four orders of magnitude in the image domain and two in video, and that no configuration dominates universally. We further reinterpret the benchmark as a decision base for adaptive codec selection, formalizing a domain-specific R-D-E router with admissible pool, oracle, and regret. A case study on the image domain, extended with a content-aware layer based on a lightweight metadata-only classifier, shows that adaptive policies reduce mean regret by up to 87% over a robust global baseline, retaining a 75.7% reduction under leave-one-dataset-out evaluation. A deployable prototype integrates hardware constraints, robust quality guards, degraded fallback, and infeasibility handling.
La valutazione dei codec multimediali è tradizionalmente fondata sul compromesso Rate–Distortion, che lega bitrate e qualità ricostruttiva. Tale paradigma, pur necessario, trascura il costo energetico operativo della pipeline di codifica e ricostruzione, dimensione che diventa centrale nel confronto tra codec classici e neurali, dove i secondi raggiungono regimi di bitrate aggressivi al prezzo di un costo computazionale superiore. Questa tesi propone una valutazione Rate–Distortion–Energy (R-D-E), in cui l'energia operativa è trattata come terzo obiettivo accanto a rate e distorsione. Il benchmark è applicato sistematicamente a tre domini: immagine (nove codec o famiglie su Kodak, Tecnick, DIV2K e CLIC2020), audio (Opus, EnCodec, DAC, SNAC, WavTokenizer su parlato, segnali ambientali e musica) e video (x264, x265, SVT-AV1, VVenC e quattro varianti DCVC sul dataset UVG in profilo LDP). Le misure energetiche sono ottenute con un protocollo a batch che integra telemetria CPU (RAPL) e GPU (NVML/Zeus), separa il costo della pipeline da quello del calcolo delle metriche percettive, e sottrae la potenza idle del sistema. I risultati mostrano che l'introduzione dell'energia modifica in modo sostanziale l'interpretazione delle prestazioni, con dinamiche energetiche superiori a quattro ordini di grandezza nel dominio immagine e a due nel dominio video, e che nessuna configurazione domina universalmente. La tesi reinterpreta inoltre il benchmark come base decisionale per la selezione adattiva dei codec, formalizzando un router R-D-E dominio-specifico con pool ammissibile, oracolo e regret. Un caso di studio sul dominio immagine, esteso con una componente content-aware basata su classifier metadata-only, mostra che policy adattive riducono il regret medio fino al 87% rispetto a una baseline globale robusta, mantenendo una riduzione del 75.7% nel protocollo leave-one-dataset-out. Il framework è accompagnato da un prototipo software eseguibile che integra vincoli hardware, soglie di qualità robuste, fallback degradato e gestione di richieste non fattibili.
Compressione dei Dati: Un'Analisi Comparativa tra Metodi Tradizionali e Approcci basati su Intelligenza Artificiale
TESIC, VALENTINO
2025/2026
Abstract
Multimedia codec evaluation has traditionally relied on the Rate–Distortion trade-off, which captures the relationship between bitrate and reconstruction quality but ignores the operational energy cost of encoding and decoding. This dimension becomes critical when comparing classical and neural codecs, as the latter often reach aggressive bitrate regimes at substantially higher computational cost. This work introduces a Rate–Distortion–Energy (R-D-E) evaluation framework that treats operational energy as a third explicit objective. The benchmark is applied across three domains: image coding (nine codecs or families across Kodak, Tecnick, DIV2K, and CLIC2020), audio coding (Opus, EnCodec, DAC, SNAC, WavTokenizer over speech, ambient sound, and music), and video coding (x264, x265, SVT-AV1, VVenC, and four DCVC variants on UVG in LDP profile). Energy measurements follow a batched protocol that combines CPU (RAPL) and GPU (NVML/Zeus) telemetry, isolates the codec pipeline from perceptual metric computation, and subtracts system idle power. Results show that incorporating energy substantially reshapes codec rankings, with energy dynamics spanning over four orders of magnitude in the image domain and two in video, and that no configuration dominates universally. We further reinterpret the benchmark as a decision base for adaptive codec selection, formalizing a domain-specific R-D-E router with admissible pool, oracle, and regret. A case study on the image domain, extended with a content-aware layer based on a lightweight metadata-only classifier, shows that adaptive policies reduce mean regret by up to 87% over a robust global baseline, retaining a 75.7% reduction under leave-one-dataset-out evaluation. A deployable prototype integrates hardware constraints, robust quality guards, degraded fallback, and infeasibility handling.| File | Dimensione | Formato | |
|---|---|---|---|
|
Tesic_Valentino.pdf
accesso aperto
Dimensione
1.63 MB
Formato
Adobe PDF
|
1.63 MB | Adobe PDF | Visualizza/Apri |
The text of this website © Università degli studi di Padova. Full Text are published under a non-exclusive license. Metadata are under a CC0 License
https://hdl.handle.net/20.500.12608/111173