Every Packet Counts: Dispersing Information for Loss-Resilient Learned Image Compression

Technical Summary: Every Packet Counts: Dispersing Information for Loss-Resilient Learned Image Compression

1. Problem Statement

Learned Image Compression (LIC) has achieved state-of-the-art rate-distortion performance but remains highly vulnerable to packet loss, a critical issue in satellite, emergency, and long-distance communications. Existing LIC methods fail in lossy environments due to two primary bottlenecks:

• Non-uniform Information Distribution: Current packetization strategies often concentrate critical information into a small subset of packets. The loss of these specific packets causes catastrophic reconstruction failure, even if subsequent packets are received intact.
• Sequential Decoding Dependencies: Entropy coding, particularly autoregressive models, creates long dependency chains where the decoding of a bitstream segment relies on all preceding bits. A single lost packet can render the remainder of the stream undecodable or cause cascading estimation errors that degrade quality significantly.

Recent attempts to address these issues, such as LossResilientLIC and ResiComp, have shown limitations. LossResilientLIC concentrates critical data in leading packets via tail-drop training, failing to solve the packetization vulnerability. ResiComp relies on hierarchical autoregressive dependencies at low bitrates, where packet loss at higher layers propagates errors downward.

2. Methodology

The authors propose an end-to-end loss-resilient compression framework designed to disperse information and minimize the impact of packet loss. The architecture, illustrated in Figure 2 of the paper, integrates three core components:

A. Inter-Channel Redistribution (ICR) and Inverse ICR (Inv-ICR)
To prevent critical information from concentrating in specific channels, the authors introduce an ICR module at the encoder. This module utilizes channel-based attention, group rearrangement, and adaptive fusion to redistribute channel energy across the latent representation ($\hat{y}$). This ensures a more uniform energy distribution before packetization. At the decoder, the Inv-ICR module reverses these operations to restore the original channel arrangement, enabling stable reconstruction even if specific channels (and thus packets) are lost.

B. Interleaved Channel Grouping (ICG)
To comply with practical packet size constraints (e.g., 900 or 1500 bytes) and ensure balanced packet importance, the latent channels are partitioned using a strided, interleaved strategy.

• The latent space is split into slices, and channels are grouped by index (e.g., even vs. odd indices) to form independent groups ($y_1, y_2, y_3, y_4$).
• Each group is further partitioned into packets based on bitrate constraints.
• This strategy ensures that every packet contains a comparable mix of information, preventing any single packet loss from causing disproportionate quality degradation.

C. Two-Layer Dual-Branch Autoregressive Structure
To mitigate the cascading effects of sequential dependencies without sacrificing entropy modeling capacity, the authors adopt a two-layer dual-branch architecture:

• Layer 1: Contains the primary information ($y_1, y_2$) and is encoded independently.
• Layer 2: Contains lower-criticality information ($y_3, y_4$) and depends on Layer 1 for distribution estimation.
• Dependency Management: The dependency chain is shortened to two levels. Crucially, the training strategy assigns low-criticality information to the second layer. Consequently, if a Layer 1 packet is lost, the resulting estimation errors in Layer 2 have minimal impact on the overall reconstruction quality.

D. Training Strategy
The model is trained using a structured masking strategy that simulates packet-level loss. Unlike previous methods that simulate loss at the channel level, this approach applies a binary mask to zero out all channels within a lost packet, accurately reflecting real-world transmission behavior. Additionally, masking in the first layer triggers masking of dependent entries in the second layer to simulate autoregressive dependency failures. The model is optimized using a rate-distortion objective ($L = R + \lambda \cdot D$).

3. Key Contributions

The paper outlines three primary contributions:

1. Information Dispersal Mechanisms: The design of the ICR mechanism and ICG strategy effectively redistributes channel energy and packet information, rendering reconstruction quality insensitive to the loss of any specific packet.
2. Robust Autoregressive Architecture: The adoption of a two-layer dual-branch structure eliminates first-layer intra-stream dependencies and limits cross-layer cascading effects, enabling robust entropy decoding under packet loss.
3. State-of-the-Art Performance: Extensive experiments demonstrate superior loss-resilient performance across multiple packet loss rates and transmission environments.

4. Experimental Results

The method was evaluated on the Kodak and CLIC datasets under uniform packet loss (5%, 10%, 20%) and bursty loss modeled by the Gilbert–Elliott (GE) channel.

• Uniform Loss Performance: At a 20% packet loss rate, the proposed method achieves an average PSNR gain of 1.84 dB over LossResilientLIC. More significantly, it reduces the PSNR variance by an order of magnitude (e.g., variance < 0.012 compared to >0.1 for competitors), indicating exceptional stability regardless of which packets are lost.
• Bitrate Efficiency: The method achieves higher PSNR at lower bitrates compared to ResiComp and LossResilientLIC. For instance, at medium bitrates with 10% loss, it outperforms ResiComp by 0.33 dB while consuming less bandwidth.
• Generalization to Bursty Loss: Notably, the model was trained only under uniform random loss. Despite this, it generalizes effectively to bursty loss conditions modeled by the GE channel, outperforming methods explicitly trained for such conditions (e.g., LossResilientLIC). This suggests that the balanced information distribution achieved by ICR and ICG provides inherent robustness to complex loss patterns.
• Visual Quality: Visual comparisons show that while competing methods produce gray artifacts or corrupted textures when early packets are lost, the proposed method yields visually clean reconstructions with minimal distortion.

5. Significance and Claims

The paper claims that its primary significance lies in addressing the dual challenges of information concentration and sequential dependency in learned image compression. By dispersing information across packets and shortening dependency chains, the proposed scheme achieves a superior balance between rate-distortion performance and loss resilience.

The authors emphasize that their approach does not require explicit training on specific bursty loss models to achieve robustness, attributing this generalization capability to the fundamental design of information dispersal. They acknowledge a shared limitation with hyperprior-based LIC approaches: the hyperprior bitstream must be received intact. However, they note that protecting this small portion of the stream (less than 8% of total bitrate) via standard Forward Error Correction (FEC) is a practical and low-overhead solution.

In conclusion, the paper posits that "Every Packet Counts" by ensuring no single packet holds disproportionate criticality, thereby transforming learned image compression into a viable solution for unreliable communication channels.