Abstract This study presents a complexity aware evaluation framework for comparing classical and neural image codecs across diverse content regimes. As video compression becomes critical for resource constrained edge IoT devices, conventional benchmarking often relies on aggregate metrics that obscure systematic performance variations driven by content complexity. Utilizing the Google Open Images Dataset (V7), we stratified approximately 125,000 images into three balanced complexity bins low,... This study presents a complexity aware evaluation framework for comparing classical and neural image codecs across diverse content regimes. As video compression becomes critical for resource constrained edge IoT devices, conventional benchmarking often relies on aggregate metrics that obscure systematic performance variations driven by content complexity. Utilizing the Google Open Images Dataset (V7), we stratified approximately 125,000 images into three balanced complexity bins low, medium, and high using original file size as a proxy for spatial entropy via quantile based thresholds. We evaluated classical anchors (JPEG, AV1, VVC) and a representative neural codec, computing BD-Rate and BD-PSNR metrics separately for each bin to ensure statistically stable estimation. Empirical results reveal that neural codecs achieve substantial bitrate savings on low-complexity content (- 32% versus JPEG), but advantages diminish significantly in high complexity regimes (-11%). Conversely, traditional codecs exhibit more robust performance across varying entropy levels. Statistical validation confirms these trends are not sampling artifacts. These findings establish content complexity as a first order variable governing codec behavior, indicating that current neural architectures struggle with high frequency details compared to hand engineered transforms. Consequently, we advocate for complexity stratified evaluation protocols as standard practice to prevent misleading performance assessments. The study highlights the necessity for complexity adaptive neural architectures and dynamic codec selection in green multimedia systems to maximize energy efficiency in heterogeneous edge environments. Future research must prioritize content conditioned rate allocation to bridge the gap between probabilistic bitstreams and dynamic network conditions. Read More Read Less

Abstract This study investigates the rate distortion complexity tradeoffs of modern neural image codecs, with emphasis on practical deployment in resource constrained environments such as edge and augmented reality devices. While neural compression models often surpass classical standards (e.g., HEVC, VVC) in rate distortion performance, their high decoding complexity particularly from autoregressive entropy models hinders real world adoption. The authors address this by evaluating three... This study investigates the rate distortion complexity tradeoffs of modern neural image codecs, with emphasis on practical deployment in resource constrained environments such as edge and augmented reality devices. While neural compression models often surpass classical standards (e.g., HEVC, VVC) in rate distortion performance, their high decoding complexity particularly from autoregressive entropy models hinders real world adoption. The authors address this by evaluating three representative architectures on the Kodak dataset (24 natural RGB images, 768×512): a hyperprior baseline, an autoregressive context model, and a transformer based codec. To ensure robust analysis, images were objectively stratified into low , medium, and high complexity bins based on Sobel-based gradient energy. Results demonstrate that transformer based codecs achieve approximately 44% improvement in BD rate over the hyperprior baseline, whereas autoregressive models yield approximately 30% savings. Critically, these gains remain consistent across all complexity levels (variation <1.5 percentage points), indicating architectural robustness rather than content specific optimization. At 0.62 bits per pixel, transformers deliver a 2.5 dB PSNR advantage with visibly superior texture and edge preservation. All performance differences were statistically significant (p < 0.001). The findings underscore a paradigm shift from pure rate distortion optimization toward balanced rate distortion complexity design. Transformer architectures, with their capacity for global context modeling, emerge as particularly promising for next generation standards where bandwidth efficiency and visual fidelity must coexist with computational constraints. The study establishes a reproducible evaluation framework grounded in objective complexity metrics and rigorous statistical validation, offering a methodological foundation for future codec development and benchmarking. Read More Read Less

Abstract This study challenges the conventional rate distortion (RD) paradigm for image codec evaluation by demonstrating its inadequacy in resource constrained, real world deployments. We introduce a unified rate distortion complexity (RDC) framework that explicitly incorporates decoding complexity as a first class evaluation dimension alongside bitrate and reconstruction quality. Through systematic analysis across classical (JPEG, BPG, VVC-intra), learned neural (hyperprior, autoregressive), and implicit neural representation (INR)... This study challenges the conventional rate distortion (RD) paradigm for image codec evaluation by demonstrating its inadequacy in resource constrained, real world deployments. We introduce a unified rate distortion complexity (RDC) framework that explicitly incorporates decoding complexity as a first class evaluation dimension alongside bitrate and reconstruction quality. Through systematic analysis across classical (JPEG, BPG, VVC-intra), learned neural (hyperprior, autoregressive), and implicit neural representation (INR) codecs on a complexity stratified dataset, we reveal critical trade offs invisible to RD only assessment. Our results show that autoregressive neural codecs, while achieving state of the art RD performance, incur prohibitive decoding latency due to inherently sequential entropy decoding rendering them impractical for edge devices. Conversely, INR based approaches emerge as Pareto optimal under tight computational budgets (<5 ms), delivering competitive reconstruction quality with orders of magnitude lower decoding latency and linear complexity scaling with resolution unlike autoregressive methods' superlinear growth. Pareto front analysis confirms that RD optimal operating points are frequently dominated in RDC space once complexity constraints are imposed. Complexity constrained evaluations further demonstrate that codec superiority is budget dependent: INR dominates at low latency budgets, hyperprior models become competitive at moderate budgets (5-20 ms), and only at high budgets (>20 ms) do autoregressive methods justify their computational cost. These findings advocate for a paradigm shift toward RDC aware codec design and evaluation, positioning INR-based compression as a compelling solution for energy efficient, low latency streaming on heterogeneous edge devices without sacrificing perceptual fidelity. Read More Read Less