JoNet: Joint Models in Vision, Radar & Wireless

• JoNet is a term representing diverse joint models that share a common backbone while incorporating domain-specific modules for tasks like object detection, target recognition, and network coding.
• In computer vision, JoNet integrates joint salient and camouflaged object detection using an asymmetric Vision Transformer with SCJoint and SBSS to balance training and decouple decoding.
• In radar and wireless domains, JoNet encompasses frameworks for joint target and jamming recognition or joint network coding, leveraging statistical decoupling and cooperative strategies to enhance performance.

JoNet is not a single standardized construct across the literature. In current arXiv usage, it most directly denotes a generalist segmentation network for joint salient object detection (SOD) and camouflaged object detection (COD), built from the SCJoint learning scheme and the Saliency-Based Sampling Strategy (SBSS) (Hao et al., 8 Aug 2025). A separate radar paper uses the label “JointHRRP-Net (JoNet)” for joint target and jamming recognition from mixed high-resolution range profiles (HRRPs) under composite jamming (Zhao et al., 19 May 2026). In earlier wireless networking work, the term itself does not appear; the corresponding construct is “Joint Network Coding,” implemented as the Joint Network Coding–Cooperative Retransmission (JNC-CR) scheme for interfering wireless multicast groups (Qureshi et al., 2011).

1. Terminological scope and domain-specific usage

The name “JoNet” therefore denotes different “joint” models in different subfields rather than a canonical architecture. In computer vision it names a dual-task segmentation network; in radar ATR it abbreviates JointHRRP-Net; and in the supplied wireless-networking source the closest match is not a model name but a coding strategy centered on Joint Network Coding. This suggests that the term functions as a compact label for joint optimization or joint inference, with semantics determined by the host domain.

Usage	Domain	Description
JoNet	SOD and COD	Generalist segmentation network using SCJoint and SBSS
JointHRRP-Net (JoNet)	HRRP radar recognition	Joint target and jamming recognition under composite jamming
Joint Network Coding / JNC-CR	Wireless multicast	Closest mapping; the paper does not use the term “JoNet”

2. JoNet in computer vision: joint salient and camouflaged object detection

In the vision literature, JoNet is defined for two class-agnostic binary segmentation problems that share the same input–output format but encode opposite visual concepts. SOD maps an RGB image to a foreground mask for the most prominent object or objects, emphasizing high contrast, distinct boundaries, and strong semantic cues. COD also maps RGB to a binary mask, but seeks objects that blend into the background, emphasizing subtle texture, low-contrast boundaries, and fine-grained local differences. The paper characterizes these as “contradictory attributes,” and notes that prior work often concluded that joint learning degrades one or both tasks because of confusion in the shared representation and training-set imbalance, with SOD typically having about $2.5\times$ more training images than COD (Hao et al., 8 Aug 2025).

JoNet adopts the simple asymmetric Vision Transformer encoder–decoder backbone from SENet, retaining only the backbone with two ViTs and removing SENet’s LICM module. Tokens are produced by splitting the image into non-overlapping patches. The encoder maps image tokens to latent tokens via self-attention, and the decoder reconstructs a binary mask from latent tokens in an MAE-like image-to-token, token-to-image pipeline. The architectural claim of the model is that nearly all parameters can be shared across SOD and COD if the decoder is allowed to learn task-specific distribution characteristics.

This is implemented through “distribution learning modules” (DLMs) inserted into each of the 8 Transformer blocks of the decoder. Each block contains two DLMs, one for SOD and one for COD. The shared encoder and most of the decoder are common to both tasks; only these small task-specific modules differ by mode. As a result, a single input image can be processed twice at inference, once with the SOD DLMs and once with the COD DLMs, producing both a saliency map and a camouflage map.

3. SCJoint and SBSS: distribution-specific decoding and balanced joint training

The core learning scheme is SCJoint. Its key assumption is that the encoding process for SOD and COD is unified, while the decoding process should have different distribution characteristics. The paper defines the “distribution characteristics of the process” as the impact the process has on the data distribution. Each DLM contains two learnable vectors that represent the mean and variance for the features processed by a decoder block. These statistics are optimized separately by SOD and COD supervision, with the stated purpose of decoupling contradictory decoding characteristics at minimal parameter cost. The operation is described as reminiscent of LayerNorm, but not as a global training stabilizer; the authors report that DLM is superior to replacing it with LayerNorm in this role (Hao et al., 8 Aug 2025).

The optimization setup uses balanced SOD and COD subsets. The paper defines the SOD training set as $\mathcal{D}_S = \{(x_i,y_i)\}_{i=1}^{N_S}$ with $N_S = 10553$, and the COD training set as $\mathcal{D}_C = \{(x_i,y_i)\}_{i=1}^{N_C}$ with $N_C = 4040$. Training alternates one SOD mini-batch and one COD mini-batch, applies the corresponding DLM mode, and minimizes the sum of the two losses with respect to the shared parameters $\theta$ and the task-specific DLM parameters. For both tasks the loss is the Pixel Position Aware loss:

$$\mathcal{L}(x,y)=L_{BCE}^{\omega}(x,y)+L_{IOU}^{\omega}(x,y).$$

SBSS addresses the data-imbalance and data-quality problem in SOD. Using a pretrained SOD network $f$, it predicts $\hat y_i=f(x_i)$ for each SOD pair and ranks each pair by

$S = S_{\alpha} + E_{\phi} + F_{\beta} + (1 - M).$

The top $\mathcal{D}_S = \{(x_i,y_i)\}_{i=1}^{N_S}$0 SOD pairs are retained, producing $\mathcal{D}_S = \{(x_i,y_i)\}_{i=1}^{N_S}$1 of size 4040. The paper attributes three effects to this procedure: it balances data volume, filters lower-quality SOD pairs such as duplicated images with different ground truths or insufficient saliency, and reduces training time to almost half of independent training because the SOD training set is cut from 10553 to 4040.

Parameter efficiency is a central design claim. SCJoint adds only the four learnable vectors shown in the DLM schematic, amounting to about $\mathcal{D}_S = \{(x_i,y_i)\}_{i=1}^{N_S}$2K parameters in the reported setting. This is presented as negligible relative to the shared network and contrasted with encoder-sharing and decoder-sharing designs that add about $\mathcal{D}_S = \{(x_i,y_i)\}_{i=1}^{N_S}$3 and about $\mathcal{D}_S = \{(x_i,y_i)\}_{i=1}^{N_S}$4 parameters, respectively.

4. Empirical behavior of vision JoNet

The ablation study is organized around four regimes: independent training (IT), naive joint training (JT), JT with SBSS, JT with SCJoint, and the full JoNet system combining SBSS and SCJoint. The reported pattern is consistent across representative SOD and COD benchmarks. On DUTS, IT yields $\mathcal{D}_S = \{(x_i,y_i)\}_{i=1}^{N_S}$5, $\mathcal{D}_S = \{(x_i,y_i)\}_{i=1}^{N_S}$6, $\mathcal{D}_S = \{(x_i,y_i)\}_{i=1}^{N_S}$7, $\mathcal{D}_S = \{(x_i,y_i)\}_{i=1}^{N_S}$8, while naive JT improves SOD to $\mathcal{D}_S = \{(x_i,y_i)\}_{i=1}^{N_S}$9. On CAMO, however, naive JT degrades COD from IT’s $N_S = 10553$0 to $N_S = 10553$1, and on COD10K from $N_S = 10553$2 to $N_S = 10553$3. Adding SBSS partly restores COD, but the paper argues that confusion remains. Adding SCJoint removes the degradation, and the full JoNet system further improves both tasks, reaching DUTS $N_S = 10553$4, PASCAL-S $N_S = 10553$5, CAMO $N_S = 10553$6, and COD10K $N_S = 10553$7 (Hao et al., 8 Aug 2025).

A second ablation concerns where the DLMs are inserted. Decoder-only DLMs are the default and best-performing choice. Encoder-only DLMs severely harm SOD, giving DUTS $N_S = 10553$8, and encoder-plus-decoder DLMs perform worse still at $N_S = 10553$9. The paper also reports that increasing the number of decoder blocks equipped with DLMs improves performance monotonically, with decoder-all best. This is used to support the stated hypothesis that encoding is largely shareable whereas decoding should be task-conditional.

Sampling strategy matters as well. With SBSS selection fixed to Top-K, DUTS/CAMO/COD10K scores are $\mathcal{D}_C = \{(x_i,y_i)\}_{i=1}^{N_C}$0, $\mathcal{D}_C = \{(x_i,y_i)\}_{i=1}^{N_C}$1, and $\mathcal{D}_C = \{(x_i,y_i)\}_{i=1}^{N_C}$2. Bottom-K and Random-K are both worse. A separate balance study, using naive joint training with fixed COD size 4040 and varying SOD size, shows that as the SOD proportion rises from roughly $\mathcal{D}_C = \{(x_i,y_i)\}_{i=1}^{N_C}$3 to $\mathcal{D}_C = \{(x_i,y_i)\}_{i=1}^{N_C}$4, SOD performance improves while COD performance degrades, which the paper presents as direct evidence for imbalance dominance.

Against prior methods, the paper reports that JoNet achieves 14 best and 4 second-best results on five SOD datasets, and 12 best and 4 second-best results on four COD datasets. Representative SOD results include DUTS $\mathcal{D}_C = \{(x_i,y_i)\}_{i=1}^{N_C}$5, HKU-IS $\mathcal{D}_C = \{(x_i,y_i)\}_{i=1}^{N_C}$6, ECSSD $\mathcal{D}_C = \{(x_i,y_i)\}_{i=1}^{N_C}$7, and PASCAL-S $\mathcal{D}_C = \{(x_i,y_i)\}_{i=1}^{N_C}$8. Representative COD results include CAMO $\mathcal{D}_C = \{(x_i,y_i)\}_{i=1}^{N_C}$9, CHAMELEON $N_C = 4040$0, COD10K $N_C = 4040$1, and NC4K $N_C = 4040$2. The paper further reports that applying SCJoint to a third binary segmentation task, shadow detection, improves all three tasks relative to both independent training and naive three-task joint training.

The training configuration is fully specified: MAE-pretrained weights for encoder and decoder, input resolution $N_C = 4040$3, random horizontal flipping, Adam, initial learning rate $N_C = 4040$4, poly schedule with power $N_C = 4040$5, batch size $N_C = 4040$6, training for 25 epochs on $N_C = 4040$7 GPUs. Code is available at the reported GitHub repository. Failure cases remain in extremely low-contrast camouflage, heavy clutter with multiple salient targets, incomplete segmentation, and blurred boundaries. The paper interprets these as motivation for stronger global scene understanding and finer boundary modeling.

5. JointHRRP-Net (JoNet): radar recognition under composite jamming

In radar ATR, “JoNet” refers to JointHRRP-Net, a unified framework for joint target and jamming recognition from a mixed HRRP $N_C = 4040$8, with $N_C = 4040$9 after zero padding. The problem setting starts from a received baseband signal containing target echo, suppression jamming, deception jamming, and thermal noise:

$\theta$0

After pulse compression with matched filter $\theta$1, the HRRP-domain signal becomes

$\theta$2

so target-related peaks are superposed with suppression and deception components. The stated goal is to predict a one-hot target label $\theta$3 and a multi-hot jamming label $\theta$4 from this mixed HRRP (Zhao et al., 19 May 2026).

The architecture has four components. First, a statistically constrained decoupling module extracts target-dominant and jamming-dominant latent branches from the mixed representation. A shallow 1-D convolutional stem produces local features, a shared encoder compresses them, and two non-shared projections form $\theta$5 and $\theta$6. Second, correlation-guided statistical constraints compute channel-wise covariance ratios against the mixture proxy $\theta$7, then generate soft masks

$\theta$8

which gate the branches to suppress redundant cross-branch information. Third, a multi-scale temporal encoding module uses micro and macro convolutional stems, a residual SE fusion block, and a diagonal structured state-space layer (S4D) to capture both local scattering structures and long-range range-cell dependencies. Fourth, a dual-expert decision module applies a softmax target head for single-label classification and a sigmoid jamming head for multi-label recognition.

Training is organized in three stages: decoupling and reconstruction, target expert training, and jamming expert training. Reconstruction uses an SI-SNR objective for the target and jamming waveforms. Classification uses single-label cross-entropy for the target expert and binary cross-entropy with logits for the jamming expert. The paper also states that end-to-end fine-tuning yields comparable closed-set performance, and gives the combined objective

$\theta$9

as an optional formulation.

The reported datasets comprise a 12-class X-band aircraft electromagnetic simulation dataset and a 3-class measured HRRP dataset. Composite jamming samples are synthesized from the four jamming modes CI, ISRJ, SMSP, and NCJ, with non-empty subsets uniformly selected so that 1–4 simultaneous jamming types are possible. For the simulation dataset, $$\mathcal{L}(x,y)=L_{BCE}^{\omega}(x,y)+L_{IOU}^{\omega}(x,y).$$0 dB and $$\mathcal{L}(x,y)=L_{BCE}^{\omega}(x,y)+L_{IOU}^{\omega}(x,y).$$1 dB; for the measured dataset, $$\mathcal{L}(x,y)=L_{BCE}^{\omega}(x,y)+L_{IOU}^{\omega}(x,y).$$2 dB and $$\mathcal{L}(x,y)=L_{BCE}^{\omega}(x,y)+L_{IOU}^{\omega}(x,y).$$3 dB. Training uses a 75%/15%/10% train/validation/test split, dynamic synthesis each epoch for training, fixed samples for validation and test, PyTorch 2.8, AMP, AdamW, cosine annealing from $$\mathcal{L}(x,y)=L_{BCE}^{\omega}(x,y)+L_{IOU}^{\omega}(x,y).$$4 to $$\mathcal{L}(x,y)=L_{BCE}^{\omega}(x,y)+L_{IOU}^{\omega}(x,y).$$5, 100 epochs, early stopping patience 20, and seed 42.

Performance claims are strong and quantitative. On 12-class simulation target recognition, average precision, recall, and macro F1 are 93.96%, more than 7 percentage points above the strongest baseline at 86.34%. On 4-mode multi-label jamming recognition, subset accuracy is 92.97%, element-wise accuracy 97.74%, precision 97.74%, recall 98.68%, and F1 98.20%. On the measured 3-class dataset, the averaged precision, recall, and F1 are approximately 96.34%. Open-set evaluation using OpenMax reports known-class accuracy 96.29%, known-class acceptance 99.75%, and AUROC 90.53%. Ablations further show the role of the decoupling and S4D components: without statistical correlation filtering, target OA drops from 93.96% to 91.99%; without S4D it drops to 83.42%; and replacing the decoupling module with a simple autoencoder yields much poorer separation, with target cosine similarity 44.26% and leakage 0.765.

6. Joint Network Coding and the nonstandard “JoNet” association in wireless multicast

The 2011 wireless paper does not use the term “JoNet.” The relevant construct is “Joint Network Coding,” realized as the Joint Network Coding based Cooperative Retransmission scheme, JNC-CR, for two interfering wireless multicast groups (Qureshi et al., 2011). The setting contains two transmitters, AP1 and AP2, with multicast receiver sets $$\mathcal{L}(x,y)=L_{BCE}^{\omega}(x,y)+L_{IOU}^{\omega}(x,y).$$6 and $$\mathcal{L}(x,y)=L_{BCE}^{\omega}(x,y)+L_{IOU}^{\omega}(x,y).$$7, group sizes $$\mathcal{L}(x,y)=L_{BCE}^{\omega}(x,y)+L_{IOU}^{\omega}(x,y).$$8 and $$\mathcal{L}(x,y)=L_{BCE}^{\omega}(x,y)+L_{IOU}^{\omega}(x,y).$$9, and an overlap region containing $f$0 receivers in total, $f$1 from each AP, with $f$2. Each AP transmits $f$3 packets per batch, with $f$4. Packet erasures are modeled as i.i.d. Bernoulli with uniform packet loss probability $f$5, and the broadcast medium allows receivers to opportunistically overhear and cache packets from the interfering AP.

JNC-CR combines Digital Network Coding and Analog Network Coding in two layers. At the packet level, each AP forms an XOR-coded packet over $f$6 using the BENEFIT heuristic, which is deterministic and memory-based. At the physical layer, the two APs may transmit simultaneously, creating a collision whose baseband model is

$f$7

If one colliding packet is known a priori, a receiver can estimate the other by subtractive ANC decoding:

$f$8

In the overlap region, cross-group opportunistic caching supplies the side information needed for this step, after which the recovered XOR packet can be decoded into missing native packets.

The scheme operates in two stages. In the non-cooperative stage, each AP independently runs BENEFIT and retransmits its XOR-coded packet when ready; collisions may occur opportunistically. In the cooperative stage, once non-overlap receivers have recovered all losses, the APs use superimposed ACKs and shared receiver status to decide whether a simultaneous collision or a single XOR transmission has greater expected benefit. The decision rule compares

$f$9

against

$\hat y_i=f(x_i)$0

where $\hat y_i=f(x_i)$1 is the number of receivers in AP $\hat y_i=f(x_i)$2’s group for whom the coded packet is innovative, and $\hat y_i=f(x_i)$3 is the number of overlap receivers able to perform collision decoding. If $\hat y_i=f(x_i)$4 exceeds the best single-AP XOR benefit, both APs transmit simultaneously; otherwise only the AP with larger $\hat y_i=f(x_i)$5 transmits.

The paper’s analytical and simulation discussion is organized around retransmission gain. It states that JNC-CR can offer a much higher retransmission gain than either ANC, DNC, or ARQ-based retransmission, and validates this using simulation with an optimal DNC scheme. The reported qualitative outcomes are that JNC-CR consistently requires fewer retransmissions than DNC across packet loss rates $\hat y_i=f(x_i)$6; average retransmissions increase logarithmically with network size $\hat y_i=f(x_i)$7; expected transmissions per packet decrease logarithmically with larger $\hat y_i=f(x_i)$8; and JNC-CR remains better than DNC across $\hat y_i=f(x_i)$9, $S = S_{\alpha} + E_{\phi} + F_{\beta} + (1 - M).$0, $S = S_{\alpha} + E_{\phi} + F_{\beta} + (1 - M).$1, and $S = S_{\alpha} + E_{\phi} + F_{\beta} + (1 - M).$2. Practical limitations are explicit: ANC requires reasonable channel estimation of $S = S_{\alpha} + E_{\phi} + F_{\beta} + (1 - M).$3 and $S = S_{\alpha} + E_{\phi} + F_{\beta} + (1 - M).$4, timing and frequency offsets must remain bounded, receivers must cache overheard packets, and the scheme assumes reliable superimposed ACKs. The stated overhead is approximately $S = S_{\alpha} + E_{\phi} + F_{\beta} + (1 - M).$5 bits per JNC-coded packet.

Across these three usages, “JoNet” is best treated as a context-dependent label rather than a singular method. In the vision and radar papers it names concrete architectures that keep a largely shared backbone while introducing task-specific mechanisms to decouple conflicting structure: DLM-based distribution-specific decoding for SOD and COD, and covariance-guided latent separation for target and jamming. In the wireless paper, the analogous idea is not a named JoNet architecture but a layered joint strategy that turns interference itself into a decodable resource.