Dual-Statistic Synergy Operator (DSO)

• DSO is a signal decoupling and gating mechanism that leverages both channel-wise mean and peak-to-mean difference for precise feature discrimination in object detection.
• It utilizes a lightweight 1x1 convolution in the DSG module to generate adaptive channel weights, improving feature selection across multiple abstraction levels.
• Empirical validation on the MS-COCO benchmark shows that DSO consistently enhances detection accuracy with minimal computational overhead compared to alternative methods.

The Dual-Statistic Synergy Operator (DSO) is a signal decoupling and gating construct introduced for improving fine-grained feature discrimination in one-stage object detection frameworks, with primary instantiation in the YOLO-DS architecture. DSO enables explicit modeling of heterogeneous object responses across shared feature channels by synergistically leveraging two channel-wise statistics—mean and peak-to-mean difference—thus facilitating more adaptive and informative channel selection. It serves as the core for the Dual-Statistic Synergy Gating (DSG) module, providing a lightweight, plug-in mechanism that improves performance and efficiency in large-scale object detection scenarios, as validated on the MS-COCO benchmark (Huang et al., 26 Jan 2026).

1. Motivation and Conceptual Basis

Conventional one-stage detectors, such as @@@@1@@@@, process input through homogenized feature representations, resulting in competitive interference among object categories, scales, and background signals within shared channels. This "channel-wise competition" leads to suboptimal context distribution, reducing the ability to dynamically focus on content-rich or object-relevant channels. Prior solutions—SENet, CBAM, and MHSA—either operate on single statistics, pool away response structure, or incur high computational cost. SENet's reliance on global mean hinders discrimination between sparse and broad activations, while CBAM's global pooling aggregates multiple object responses. MHSA (as in Vision Transformers) mixes multiple scales within heads and is computationally intensive.

The DSO is designed to address these limitations by simultaneously distilling the channel-wise mean, $\mu_{b,c}$, and the peak-to-mean difference, $d_{b,c}$, to provide a two-dimensional cue for channel attention. This approach enables the feature processor to distinguish, for example, between a channel strongly activated by a small object (high peak, low mean) versus a large object (high mean, small peak-to-mean).

2. Mathematical Definition and Workflow

Given input tensor $x \in \mathbb{R}^{B \times C \times H \times W}$ (batch size $B$, channels $C$, spatial dimension $H \times W$), DSO computes:

• Channel-wise Mean:

$$\mu_{b,c} = \frac{1}{H \cdot W} \sum_{h=1}^H \sum_{w=1}^W x_{b,c,h,w}, \quad \mu \in \mathbb{R}^{B \times C \times 1 \times 1}$$

• Channel-wise Maximum:

$$m_{b,c} = \max_{h,w} x_{b,c,h,w}, \quad m \in \mathbb{R}^{B \times C \times 1 \times 1}$$

• Peak-to-Mean Difference:

$$d_{b,c} = m_{b,c} - \mu_{b,c}, \quad d \in \mathbb{R}^{B \times C \times 1 \times 1}$$

• Synergistic Decision Response (DSO Core):

$$y_{b,c} = \Phi(\mu_{b,c}, d_{b,c}) = (d_{b,c} + 1)(\mu_{b,c} + 1) - 1 = \mu_{b,c} \cdot d_{b,c} + \mu_{b,c} + d_{b,c}$$

yielding $y \in \mathbb{R}^{B \times C \times 1 \times 1}$.

Subsequently, a gating network consisting of a $1 \times 1$ convolution and sigmoid activation transforms this synergy response into per-channel weights:

• $$z_{DSG} = W_{DSG} * y + b_{DSG}, \quad W_{DSG} \in \mathbb{R}^{C' \times C \times 1 \times 1}, \quad b_{DSG} \in \mathbb{R}^{C'}$$ with output dimension $C' = \lfloor C/2 \rfloor \times (2 + n)$, where $n$ is the number of bottleneck blocks in YOLOv8 C2F modules (with $n=3$ as typical).
• Sigmoid gating:

$$w_{DSG} = \sigma(z_{DSG}), \quad w_{DSG} \in \mathbb{R}^{B \times C' \times 1 \times 1}$$

This gating weight is broadcast across spatial dimensions and applied to the C2F-concatenated feature $$x_{cat} \in \mathbb{R}^{B \times C' \times H \times W}$$:

$x_{out} = w_{DSG} \odot x_{cat}$

3. Integration within YOLO-DS and Architectural Considerations

The DSG module, which operationalizes DSO, is inserted at each C2F concatenation point in both backbone and detection head of YOLO-DS. After concatenation of $n$ bottleneck branches to form $x_{cat}$ in the C2F block, DSG modulates $x_{cat}$ using learned channel-wise gates derived via the DSO mechanism. This procedure is designed to adaptively gate feature channels at multiple abstraction levels, supporting more nuanced separation of small object, large object, mixed, and background features.

Parameter and computational cost breakdown (for YOLOv8-L, $C \approx 512, n = 3, C' = 1\,280$):

• Number of learnable parameters: $512 \times 1\,280 + 1\,280 = 656\,640$
• FLOPs per $1 \times 1$ conv at $640 \times 640$ spatial: $$512 \times 1\,280 \times 640 \times 640 \approx 2.7 \times 10^{11}$$ operations
• Net increase over YOLOv8-L backbone: $\sim 4.4$ GFLOPs
• Measured latency overhead (TensorRT, T4/4090): 0.25 ms (YOLOv8-L), representing a minimal inference impact (Huang et al., 26 Jan 2026).

4. Comparison with Alternative Mechanisms

Ablation and comparative studies contextualize DSG/DSO among prior channel-attention and scale-decomposition approaches:

• SENet: Processes only channel mean, rendering it insensitive to scale heterogeneity.
• CBAM: Combines mean and max, but global pooling entangles responses from multiple objects or object-background mixtures.
• MHSA (ViT): Attends across heads and scale but introduces quadratic cost in both computation and memory.

DSO is unique in that it provides a compact, per-channel 2D feature statistic ($\mu$, $d$), enabling precise object-scale and activity discrimination while incurring only linear cost and introducing an efficient $1 \times 1$ convolutional transformation for gating.

5. Empirical Validation

Experimental analysis on the MS-COCO benchmark demonstrates tangible benefits in detection accuracy with marginal resource overhead. Key quantitative results include:

Model	Baseline AP (%)	+DSG AP (%)	ΔAP	Baseline Latency (ms)	+DSG Latency (ms)	ΔLatency (ms)
YOLOv8-N	37.3	38.7	+1.4	1.47	1.51	+0.04
YOLOv8-S	44.9	46.6	+1.7	2.66	2.77	+0.11
YOLOv8-M	50.2	51.6	+1.4	5.86	6.09	+0.23
YOLOv8-L	52.9	54.1	+1.2	9.06	9.31	+0.25
YOLOv8-X	53.9	55.0	+1.1	14.37	14.99	+0.62

Fine-grained ablation (YOLOv8-L, AP=52.9% baseline): DSG alone yields AP=53.5% (+0.6), with a parameter increase to 49.8M (+6.1M) and FLOPs to 170.1 (+4.4). In aggregate with depth-wise MSG, net AP gains reach 1.1–1.7% across scales at sub-0.3 ms latency cost (Huang et al., 26 Jan 2026).

6. Implementation Overview and Pseudocode

The algorithmic flow of the DSG block, which encapsulates the DSO, is as follows:

def DSG_Block(x: Tensor[B,C,H,W], n_bottleneck: int): mu = x.mean(dim=[2,3], keepdim=True) # μ, shape [B,C,1,1] m = x.amax(dim=[2,3], keepdim=True) # m, shape [B,C,1,1] d = m - mu # d, shape [B,C,1,1] y = (d + 1) * (mu + 1) - 1 # y, shape [B,C,1,1] C_prime = floor(C / 2) * (2 + n_bottleneck) z = Conv1x1(in_channels=C, out_channels=C_prime)(y) # z, shape [B,C',1,1] w = sigmoid(z) # w, shape [B,C',1,1] x_cat = get_C2F_concatenation(x) # [B,C',H,W] x_out = w.expand_as(x_cat) * x_cat # [B,C',H,W] return x_out

Weights are shared across all spatial positions per channel, and the $1\times1$ convolution allows for learnable coupling of synergistic statistics.

7. Significance and Applicability

DSO, as operationalized in the DSG module, delivers a parameter-efficient, statistically informed solution to the challenge of heterogeneous channel competition and attention in deep convolutional networks. Its successful deployment within YOLO-DS establishes a generalizable paradigm for adaptive feature gating in multi-scale object detection contexts. Empirical evidence substantiates consistent accuracy improvements with negligible latency cost, supporting its adoption in resource-sensitive inference deployments (Huang et al., 26 Jan 2026). A plausible implication is that similar dual-statistic approaches could be considered for other attention or selection modules where channel heterogeneity is pronounced.