Selective Depthwise Separable Convolution for Lightweight Joint Source-Channel Coding in Wireless Image Transmission

arXiv:2604.22338 · eess.IV, cs.CV · Submitted 2026-04-24 · Read on arXiv

Listen

Transcript

Introduction to the show: ident: AI Radio. Generated commentary on the latest Artificial Intelligence papers.

Tom: Today's paper: "Selective Depthwise Separable Convolution for Lightweight Joint Source-Channel Coding in Wireless Image Transmission".

Jane: The gist The proposed framework incorporates a selective replacement strategy that enables flexible Conv-to-DSConv replacement at different replacement ratios and positions.

Tom: First, who's behind it and why it matters.

Paper summary: Tom: So we're looking at this paper today, "Selective Depthwise Separable Convolution for Lightweight Joint Source-Channel Coding in Wireless Image Transmission." It’s about taking deep learning based joint source-channel coding, or JSCC, and making it lighter.

Jane: That’s right. The main idea is that depthwise separable convolutional layers have worked well to cut down on how much computing power these systems need for wireless image transmission, but nobody had really looked closely at swapping out standard convolutional layers with these depthwise separable ones in JSCC systems yet.

Lu: What the paper proposes is a configurable lightweight JSCC framework called DSC-JSCC that uses a selective replacement strategy. It lets you choose where and how much to replace standard convolutional layers with their depthwise separable counterparts based on different ratios and positions.

Meng: So, if I'm getting this right, it’s not just about swapping one thing for another; it’s about having a flexible system that can adjust the compression level by changing the replacement ratio and the layer position.

Lalam: Exactly. The paper claims they can get different model complexities and then analyze how that complexity change affects how well they reconstruct the image, which is pretty interesting because you get to tune it.

Tom: Right, so what’s actually in this paper? It sets up a system where an image is mapped into a complex vector by an encoder, transmitted over a noisy wireless channel, and then decoded back into an image. The whole thing is end-to-end deep learning based JSCC <ref:2604.22338#pg1>.

Jane: The authors then introduce DSC-JSCC as this framework that puts a selective replacement strategy in place. They specifically replace Conv layers in the encoder with DSConv and transposed convolutional, or DSTConv, layers in the decoder with DSTConv layers <ref:2604.22338#pg2>.

Lu: The paper shows some specific variants too, like DSC-JSCC-sixty (E1D1), which is a representative example of this framework. It details that each DSConv layer has a depthwise convolution for spatial feature extraction followed by a pointwise convolution for channel mixing, and DSTConv layers use transposed depthwise convolution for upsampling followed by a pointwise convolution <ref:2604.22338#pg3>.

Paper summary: Meng: They also set up the training process by optimizing the encoder and decoder end-to-end using mean squared error loss during an offline training phase, before testing on real data in the online reconstruction phase, which is what equation four describes <ref:2604.22338#pg3>.

Lalam: The core contribution they highlight is that by varying the replacement ratios and positions, you can get different levels of model compression while trying to keep the reconstruction performance similar to other models like Deep-JSCC five and SwinJSCC ten.

Tom: And they found something specific about where these replacements happen. They systematically analyzed how replacing Conv layers at different positions in the encoder and decoder affects quality under a fixed replacement ratio.

Jane: Their findings pointed to the intermediate layers being key. Specifically, they showed that Conv-to-DSConv replacement at the intermediate layers of both the encoder and decoder achieves a good trade-off between complexity and performance <ref:2604.22338#pg2>.

Lu: What that means for us is that there’s some kind of layerwise redundancy in these deep learning based JSCC systems, which they found through this systematic analysis <ref:2604.22338#pg1>.

Meng: From an engineering standpoint, if you're trying to deploy something on a resource-constrained edge device, knowing that the complexity is concentrated in those intermediate layers is really helpful for designing your network architecture <ref:2604.22338#pg1>.

Lalam: They also showed that their framework can achieve a substantial parameter reduction, say fifty-four point eight percent compared to Deep-JSCC, and a FLOPs reduction of sixty-one point seven percent for DSC-JSCC-sixty (E1D1) <ref:2604.22338#pg3>.

Tom: That’s a big number, so we can see it’s actually making the models much smaller and faster without sacrificing too much quality <ref:2604.22338#pg1>. But they also tested extreme compression levels, like eighty percent replacement, and those suffered significant performance degradation <ref:2604.22338#pg1>.

Jane: So, the main lesson here is that you can find a sweet spot where you get a good balance of being light enough for deployment while still maintaining decent reconstruction quality <ref:2604.22338#pg1>.

Lu: I think the real implication is that this gives us more flexibility in designing these neural network structures, moving away from one fixed architecture to something adaptable based on the specific constraints of the wireless link and device you’re targeting <ref:2604.22338#pg2>.

Paper summary: Meng: For practical application, it means we don't have to just pick one pre-defined efficient model; we can actually tune it for our specific hardware limitations by choosing the right replacement strategy <ref:2604.22338#pg1>.

Lalam: And looking at this from a generative model view, Lalam thinks that this ability to tailor the network structure based on compression needs could make future generation of efficient source-channel models much more diverse and specialized <ref:2604.22338#pg1>.

Tom: Well, let’s talk about the title and authors for a second. The paper is titled "Selective Depthwise Separable Convolution for Lightweight Joint Source-Channel Coding in Wireless Image Transmission," written by Ming Ye, Kui Cai, Cunhua Pan, Zhen Mei, Wanting Yang, and Chunguo Li <ref:2604.22338#pg1>.

Jane: It’s a very focused piece of work that tackles the complexity reduction problem directly within the JSCC domain for wireless images <ref:2604.22338#pg1>.

Lu: The paper's main contribution is proposing DSC-JSCC, which is this configurable lightweight DLbased JSCC framework that uses a selective replacement strategy to reduce computational complexity while keeping performance comparable <ref:2604.22338#pg1>.

Meng: It really matters because it shows that we don't have to stick with one monolithic architecture when trying to make these things practical for real-world, resource-constrained edge devices <ref:2604.22338#pg1>.

Lalam: It suggests a path forward where the network structure itself becomes part of the optimization process, rather than just training a fixed network and hoping for the best <ref:2604.22338#pg1>.

Tom: So, to wrap up this segment, we’ve seen that this paper offers a systematic way to explore how layer replacement impacts complexity versus reconstruction quality in JSCC systems <ref:2604.22338#pg1>.

Jane: And the key finding is that replacing layers at the middle depths of both encoder and decoder gives you a favorable trade-off <ref:2604.22338#pg2>.

Lu: This opens up new avenues for designing efficient deep learning models tailored to specific communication constraints <ref:2604.22338#pg1>.

Meng: For deployment, it means we can design systems that are actually optimized for the hardware they run on, not just theoretical ideals <ref:2604.22338#pg1>.

Lalam: And ultimately, it gives us a tool to generate more diverse and specialized lightweight models for image transmission tasks <ref:2604.22338#pg1>.

Conclusion: Tom: So we're wrapping up this look at the paper, "Selective Depthwise Separable Convolution for Lightweight Joint Source-Channel Coding in Wireless Image Transmission."

Jane: Basically, they’ve put together a way to make deep learning joint source-channel coding models much smaller by smartly swapping out some standard convolutional layers for depthwise separable ones.

Lu: It’s about this selective replacement strategy, so you can decide at different points in the network whether to use the full convolution or the depthwise separable version based on some ratio.

Meng: The authors found that if you tune those ratios and positions right—specifically in the middle layers of both the encoder and decoder—you get a really good balance between how small it gets and how well it reconstructs images.

Lalam: It means there’s this hidden redundancy in these deep learning JSCC systems, so you don't have to use one big fixed model if you can tailor the structure.

Tom: Exactly. They show that with a sixty percent replacement ratio, they get a massive reduction in parameters and floating point operations, maybe over sixty percent less computation than the original models they compared it to.

Jane: The authors tested this on real images like CIFAR-ten and CelebA, and while the reconstruction quality doesn't just drop off a cliff when you go for that heavy compression, it stays pretty good for smaller numbers.

Lu: The finding is that DSC-JSCC-sixty which they call E2D2, actually hits the lowest complexity among all their variants with that fixed replacement ratio because the cost really concentrates in those intermediate layers.

Meng: For someone looking at deploying this on a device with limited processing power, knowing exactly where the cost lives—in the middle layers—that’s super practical info for designing the hardware side.

Lalam: This capability to tailor the network structure based on compression needs suggests that future generation of efficient source-channel models could become much more diverse and specialized for different communication links.

Tom: It moves us past just building one big, powerful model and toward a system where we can customize the architecture for specific constraints.

Jane: And this opens up a whole new way to think about how we design these neural networks for real-world wireless image transmission problems.

eess.IV, cs.CV

Submitted: 2026-04-24

Updated: 2026-10-09

Comments: 6 pages, 6 figures, journal

License: http://arxiv.org/licenses/nonexclusive-distrib/1.0/

Importance score: 69/100

The gist: The gist The proposed framework incorporates a selective replacement strategy that enables flexible Conv-to-DSConv replacement at different replacement ratios and positions.

Key concepts

Depthwise Separable Convolution (DSConv): This is a technique used to reduce the number of parameters and computations in standard convolutional layers. It splits the standard convolution into two steps: first, a depthwise convolution that applies a single filter to each input channel separately, and second, a pointwise convolution that combines these outputs. This significantly lowers computational load while maintaining feature extraction capabilities.
Joint Source-Channel Coding (JSCC): JSCC is an end-to-end deep learning system designed to compress an image source into a complex vector for transmission over a wireless channel, and then reconstruct the original image from the received noisy signal. The system involves a trainable encoder that compresses the image and a trainable decoder that reconstructs it.
Selective Replacement Strategy: This is the core methodology where researchers selectively swap standard convolutional layers (Conv) with their depthwise separable counterparts (DSConv) based on specific replacement ratios and layer positions. By varying these parameters, the framework allows for flexible model compression, enabling optimization for different levels of complexity while analyzing how layer placement affects reconstruction quality.

Terminology

Edge Devices — 1× in this paper · explained in 2 episodes
Redundancy — 1× in this paper · explained in 2 episodes

Summary

The gist The proposed framework incorporates a selective replacement strategy that enables flexible Conv-to-DSConv replacement at different replacement ratios and positions.

System Model

The system under consideration is an end-to-end DL-based JSCC transmission system consisting of a trainable encoder fθ, a trainable decoder gψ, and a noisy wireless channel as shown in Fig. 1 The source image x ∈ R d is mapped into a complex-valued vector z ∈ C k by the encoder as z = fθ (x, ρ) ∈ C k, where ρ is the bandwidth compression ratio defined as ρ = k/d After encoding, the vector z is transmitted over a Rayleigh slow-fading channel, yielding zˆ = η(z, σ2) = hz + n Finally, the corrupted vector zˆ is fed into the decoder to reconstruct the image as xˆ = gψ (zˆ) ∈ R d.

Proposed Framework

The proposed framework is termed DSC-JSCC and incorporates a selective replacement strategy that enables flexible Conv-to-DSConv replacement at different layer positions and ratios Specifically, Conv layers in the encoder are replaced with DSConv layers, while TConv layers in the decoder are replaced with DSTConv layers, thereby reducing computational complexity. The architecture of DSC-JSCC-60 (E2D2), a representative variant of the proposed framework, is illustrated in Fig. 2 Each DSConv layer consists of a depthwise convolution for spatial feature extraction followed by a pointwise convolution for channel mixing. Each DSTConv layer consists of a transposed depthwise convolution for spatial upsampling followed by a pointwise convolution for channel mixing. The framework consists of two phases: the offline training phase and the online reconstruction phase. During the offline training phase, the encoder and the decoder of the proposed framework are optimized end-to-end by minimizing the mean squared error (MSE) loss.

Layer Replacement Strategies

The framework adopts a selective replacement strategy based on replacing different ratios of Conv layers in the encoder and TConv layers in the decoder with DSConv and DSTConv layers, respectively. Varying replacement ratios leads to different levels of model compression, and the structures of these ratio-based variants are summarized in Table I. A higher value of X corresponds to a higher level of model compression, thereby further reducing computational complexity. Additionally, DSC-JSCC variants are generated by replacing Conv and TConv layers at different positions in the encoder and the decoder with DSConv and DSTConv layers, respectively. The early, middle, and late Conv layers in the encoder and TConv layers in the decoder are replaced under a 60% replacement ratio to evaluate the effect of layer positions.

Performance Evaluation

To evaluate the performance of the DSC-JSCC variants against three baseline models, namely Deep-JSCC [5], SwinJSCC (without SNR and rate adaptation) [10], and DeepJSCCT [12], we use the CIFAR-10 [16] and CelebA [17] datasets. The quality of the reconstructed images is evaluated using the peak signal-tonoise ratio (PSNR) and the learned perceptual image patch similarity (LPIPS) [18]. As shown in Fig. 3(a), the PSNR of the DSC-JSCC variants generally decreases as the replacement ratio increases. Nevertheless, both DSC-JSCC-20 and DSC-JSCC-40 achieve PSNR comparable to that of Deep-JSCC while exhibiting lower computational complexity. DSC-JSCC-60 (E1D1) reduces the number of parameters by 54.8% and the FLOPs by 61.7% compared to Deep-JSCC, thereby improving computational efficiency. Among the DSCJSCC-60 variants with a fixed replacement ratio of 60%, DSCJSCC-60 (E2D2) achieves the lowest computational complexity.

Conclusion

Overall, DSC-JSCC-60 (E2D2) achieves a favorable complexity-performance trade-off on the CelebA dataset. This is because the computational cost is mainly concentrated in the intermediate layers of the encoder and the decoder. DSC-JSCC-60 (E1D1) reduces the number of parameters by 54.8% and the FLOPs by 61.7% compared to Deep-JSCC, thereby improving computational efficiency. Among the DSCJSCC-60 variants with a fixed replacement ratio of 60%, DSCJSCC-60 (E2D2) achieves the lowest computational complexity. Although both DSC-JSCC-80 and DSCJSCC-100 achieve lower computational complexity than DSCJSCC-60 (E2D2), they suffer from significant performance degradation. Overall, DSC-JSCC-60 (E2D2) achieves a favorable complexity-performance trade-off, making it attractive for practical deployment in resource-constrained scenarios.

Improvements for AI systems

Bold header: Configurable Lightweight JSCC Framework

This framework enables flexible Conv-to-DSConv replacement at different replacement ratios and positions, allowing for controllable model complexity while maintaining comparable reconstruction performance through a selective replacement strategy.

Bold header: Layer-wise Redundancy Identification

The analysis reveals that Conv-to-DSConv replacement at the intermediate layers of the encoder and decoder achieves a favorable complexity-performance trade-off, revealing layerwise redundancy in DL-based JSCC systems. This insight guides future architectural design by suggesting where complexity reduction yields maximal benefit without significant performance loss.

Bold header: Optimized Resource Allocation for Edge Deployment

The proposed framework achieves substantial parameter reduction with only slight performance degradation, enabling flexible complexity-performance trade-offs for resource-constrained edge devices. This capability allows AI systems to be deployed effectively on hardware with limited computational, memory, and energy resources by tailoring the model size precisely to the required latency and accuracy.