Sending a photograph from an underwater robot to the surface is slow. Sound is the only practical carrier, and acoustic modems deliver somewhere between a few bytes and about 20 kilobits per second. Field trials show that the dependable mode is short 64-byte messages, and that higher-throughput bursts fall apart when the water column is layered by temperature and density. The current best compressors for this setting are large neural codecs. They assume every bit arrives, they run on workstation GPUs, and they have been tested only offline. We propose the opposite allocation of work. The robot runs a deliberately simple encoder: a block cosine transform, coarse quantization, and an interleaved packet layout that a microcontroller can execute. The receiver, which sits on a ship or shore station with a GPU, treats decoding as an optimization problem. Every coefficient that arrives fixes a narrow interval the image must respect. Every coefficient lost in a dropped packet simply removes a constraint. A smoothness prior fills in whatever the constraints leave open.
The choice of prior is what connects this to the variational imaging literature. The classic choice, total variation (a penalty on the summed magnitude of image gradients), prefers piecewise-constant images. At very low bitrates and with many coefficients missing, it turns the smooth light falloff and backscatter haze that dominate underwater scenes into flat terraces, an effect called "staircasing." Total generalized variation (TGV), a second-order generalization that tolerates linear ramps, should avoid this while keeping edges sharp. The note claims that this combination gives graceful degradation under real bursty packet loss. It pairs a cheap interleaved encoder with a TGV decoder whose feasible set shrinks as packets arrive. We also claim that the claim can only be judged honestly against packet-loss traces recorded on real modems, because idealized simulation is likely to rank the competing methods differently.
Core Claim
A Total Variation-Based JPEG Decompression Model supplies the method we reuse. Bredies and Holler decode JPEG by minimizing total variation subject to each cosine coefficient lying in its known quantization interval. They solve the problem with a GPU primal-dual algorithm that stops on the primal-dual gap, running about 10 ms per 512×512 image and 0.26 s versus over 28 s on CPU for a 4272×2848 image. They report staircasing on smooth and textured regions and name TGV as the remedy. They work only with complete bitstreams of standard optical test images.
Our decoder extends their constraint set to lossy channels. Let $u$ be the image, $A$ the block cosine transform, $R$ the set of coefficient indices whose packets arrived, and $[l_k, r_k]$ the quantization interval of received coefficient $k$. The decoder solves
$$\hat u = \arg\min_u \ \mathrm{TGV}^2_\alpha(u) \quad \text{subject to} \quad (Au)_k \in [l_k, r_k] \ \text{for all } k \in R,$$
Here $\mathrm{TGV}^2_\alpha$ is the second-order functional of Regularization of linear inverse problems withtotal generalized variation, with weights $\alpha_0, \alpha_1$ balancing edges against ramps. That paper supplies the well-posedness theory but no solver, and we provide the solver. Lost coefficients carry no constraint, or optionally a wide prior interval taken from coefficient statistics. The decoder therefore needs no special concealment logic. A dropped packet makes the problem less constrained, and the prior decides what to put there.
The novelty is not "apply TGV to underwater images." It is a specific coupling of three pieces. The first is a packet layout designed so that losses become scattered, recoverable gaps in the constraint set. The second is a second-order prior chosen because the information the packets fail to carry is exactly the smooth illumination structure that first-order total variation destroys. The third is an evaluation on recorded acoustic loss traces rather than idealized random drops.
Research Program
Encoder and packetization. The encoder computes an 8×8 block cosine transform and quantizes coarsely to a target of roughly 0.05 to 0.1 bits per pixel. That range matches the regime in which RFD-ECNet: Extreme Underwater Image Compression with Reference to Feature Dictionary operates. A 512×512 image at 0.05 bpp is about 1.6 kB, roughly 26 of the 64-byte instant messages that Developing an Underwater Network of Ocean Observation Systems with Digital Twin Prototypes - A Field Report from the Baltic Sea found reliable in the Baltic Sea. We use their 64-byte payload as the design unit.
Each packet is entropy-coded independently, so one loss cannot desynchronize later packets. The packet's sequence number determines, through a fixed pseudo-random interleaving, which blocks and frequency bands it carries. A burst of consecutive lost packets therefore removes coefficients spread thinly across the image instead of blanking a region. The lowest-frequency coefficients are carried redundantly in two packets. That redundancy is the only forward error correction we use, and its byte cost is charged against the rate.
Decoder. The decoder is a primal-dual solver for the TGV problem above, following Bredies and Holler's GPU design. Decoding is "anytime": it starts when the first packets arrive and warm-starts each time new constraints appear. On a channel that delivers a packet every second or so, the receiver has seconds of compute per update, far more than the millisecond-scale solves Bredies and Holler report. The efficiency argument of this proposal therefore sits on the sender. A transform-and-quantize encoder replaces the 86.7G-FLOP neural codec of RFD-ECNet. We will measure, not assume, the encoder's energy and time on a representative embedded processor.
TGV's two weights are an acknowledged open problem. We will fit a single simple rule mapping bitrate and observed loss fraction to the weights, tune it on a held-out split, and report sensitivity curves rather than hide the tuning.
Measuring staircasing directly. PSNR barely registers terraces. We therefore add a staircasing score: the fraction of pixels in reference-smooth regions where the decoded gradient is near zero while the reference gradient is not. This is computed on the reference images of An Underwater Image Enhancement Benchmark Dataset and Beyond. That paper also shows that the common no-reference underwater scores disagree with human judgment. For that reason we include a blinded pairwise human rating on a subset, and we do not rely on those scores.
Real loss traces. We will record packet-arrival logs from commercial acoustic modems sending 64-byte messages, in a test tank and, if access allows, in a coastal deployment. We will run identical encoder and decoder code on synthetic loss and on trace replay. This uses a shared code path in the style of the Barbie et al. environment flag, so that any difference comes from the channel and not the software.
Baselines
The naive reference point is plain JPEG-style decoding of the same packets: zero-fill lost coefficients and take the inverse transform. It shows how much any prior buys. It should fail visibly as soon as bursts remove low-frequency content.
The closest variational competitor is the total-variation decoder of A Total Variation-Based JPEG Decompression Model. We run it with exactly our constraint sets, packets, solver family, and stopping rule, and give its weight the same tuning budget. Swapping TGV for TV is the ablation that isolates our key ingredient. Beating it means equal or better PSNR together with a clearly lower staircasing score and a human preference, at matched bytes and matched loss traces. The file reports runtimes for this method but no reconstruction quality on underwater data, and no reported result found on Lacuna for TV-constrained decoding of underwater images.
A cheaper anti-staircasing ablation replaces TGV with Huber-TV, which Infimal Convolution Regularisation Functionals of BV and Lp Spaces - Part I: The Finite p Case proves equivalent to its $TVL^2$ functional. We also run its higher-$p$ variants, which preserve affine structure without second-order fields. If Huber-TV or $TVL^p$ matches TGV, we will recommend the cheaper one.
The strongest existing compressor is RFD-ECNet. It reports an average 31.7% BD-rate (bitrate saving at equal quality) over VVC-intra below 0.1 bpp and down to 0.049 bpp, at 86.7G FLOPs, on UGWI, EUVP, UFO, and UIEB. We will run it, if code is obtainable or can be reimplemented, at the same byte budgets on the same test images. We packetize its bitstream into 64-byte messages and give it the same per-packet byte allowance for forward error correction, so that error correction costs both methods the same bytes per image. On lossless channels we expect it to win in rate-distortion. Beating it means better quality under recorded loss traces at equal total bytes, at a sender cost orders of magnitude lower.
Finally, a learned post-processor runs at the receiver on zero-filled decoded images. It is trained on the training split with synthetic random erasures and gets the same GPU and the same data. This is the most dangerous competitor, discussed below.
Why not just…?
Why not just use total variation, since it is cheaper and well understood? Because the information most likely to be missing at 0.05 bpp with lost packets is the smooth background, and TV fills gaps with plateaus. G-Norm Properties of Bounded Variation Regularization characterizes this piecewise-constant bias as intrinsic to the regularizer, and Bredies and Holler observe it in JPEG decoding. Receiver compute is not scarce, so TV's speed advantage buys nothing here. The TV-versus-TGV ablation settles the question. If staircasing scores and human preferences show no difference on underwater scenes, the second-order prior is unnecessary and we will say so.
Why not just train a neural artifact-and-erasure remover at the receiver, where FLOPs are free? This is a real threat. The receiver has a GPU, so the FLOP argument against deep models does not apply on that side. Our counterargument is distribution shift. UIEB documents deep enhancers failing on unfamiliar water types, and a network trained on synthetic erasure patterns may not handle real burst structure. A learned model can also produce details that contradict the received coefficients, whereas our decoder's output provably satisfies every received interval. We will test the network on held-out water types and on recorded traces it never saw. If it still wins, the honest outcome is a hybrid: the network's output initializes our constrained solver, which projects it back onto the data-consistent set. We will report that hybrid as well.
Why not just add stronger forward error correction and decode normally? The distance-adaptive redundancy of Internode Distance-Based Redundancy Reliable Transport in Underwater Sensor Networks avoids retransmission round trips. Every redundant byte, however, is a byte not spent on image content. A reconstruction-aware decoder may make much of that redundancy unnecessary. We compare at equal total bytes, sweeping the redundancy fraction for the plain decoder. If heavy error correction plus plain decoding matches our method on real traces, our contribution reduces to the prior on clean channels.
Why not just evaluate with an acoustic channel simulator? We do use synthetic loss for sweeps. But Barbie et al. found burst mode collapsing under thermoclines in ways simulation did not reveal. Polar Transversal Initial Alignment Algorithm for UUV with a Large Misalignment Angle reports a simulated roll-axis error of 92.04′ against 208.4′ in a semi-physical pool trial. The decisive check below tests directly whether simulation changes the conclusions.
Proposed Experiments
Test images come from UIEB's 890 reference images, its 60 challenging images, and from UGWI (2,150 images), EUVP, and UFO. Water types are held out across splits for tuning and testing. We sweep target rates from 0.05 to 0.1 bpp and loss conditions of none, independent random drops at several rates, a two-state bursty model, and recorded modem traces. Each method is reported with PSNR, SSIM, the staircasing score, blinded pairwise human preference, measured encoder time and energy on embedded hardware, and decoder wall-clock to a fixed primal-dual gap.
The first decisive check is the TV-to-TGV swap under identical constraints. The proposal's central mechanism stands or falls on whether TGV lowers the staircasing score and wins human preference in smooth-water regions without losing edges. The second is whether interleaved packetization beats contiguous block-ordered packetization under recorded bursts; without that, the constraint-relaxation story does not hold. The third is a rank-order comparison of all methods under synthetic loss versus trace replay. If the ordering of TGV, TV, the learned post-processor, and RFD-ECNet changes between the two, that change is itself a reportable result about evaluating underwater image links in simulation. If it does not change, cheaper simulated evaluation is vindicated for this problem.
References
- A Total Variation-Based JPEG Decompression Model
- Regularization of linear inverse problems withtotal generalized variation
- RFD-ECNet: Extreme Underwater Image Compression with Reference to Feature Dictionary
- Developing an Underwater Network of Ocean Observation Systems with Digital Twin Prototypes - A Field Report from the Baltic Sea
- An Underwater Image Enhancement Benchmark Dataset and Beyond
- Infimal Convolution Regularisation Functionals of BV and Lp Spaces - Part I: The Finite p Case
- G-Norm Properties of Bounded Variation Regularization
- Internode Distance-Based Redundancy Reliable Transport in Underwater Sensor Networks
- Polar Transversal Initial Alignment Algorithm for UUV with a Large Misalignment Angle