---
title: 'DRFSPPF: Dilated Receptive Field SPPF Module'
url: https://www.emergentmind.com/topics/dilated-receptive-field-sppf
type: topic
---

# DRFSPPF: Dilated Receptive Field SPPF Module

Searching arXiv for the specified paper and closely related context papers on SPPF and dilated receptive-field modules.
Dilated Receptive Field SPPF, abbreviated DRFSPPF, is a backbone-end feature aggregation block introduced in "Cott-ADNet: Lightweight Real-Time Cotton Boll and Flower Detection Under Field Conditions" [2509.12442]. It replaces the standard SPPF at the end of the YOLOv11n backbone and is designed to expand receptive fields for more effective multi-scale context modeling at low computational cost. Within Cott-ADNet, the module is presented as a mechanism for enhancing spatial representation and robustness under complex field conditions, especially where cotton bolls and flowers exhibit weak or low-contrast features. Its defining characteristics are multi-scale max-pooling, large-kernel directional depthwise convolution, dilated depthwise convolution, residual recalibration, and channel projection, all arranged so that the output retains the same spatial dimensions and channel count as the backbone feature map [2509.12442].

## 1. Position within the Cott-ADNet architecture

DRFSPPF is introduced as one of two new modules in Cott-ADNet, a lightweight real-time detector built on YOLOv11n for cotton boll and flower recognition under field conditions [2509.12442]. The module directly replaces the vanilla SPPF at the end of the backbone. Its input is the final backbone feature map, denoted \(F \in \mathbb{R}^{H \times W \times C}\), and its output remains \(H \times W \times C\), so it can be routed into the PANet/fusion neck and subsequently to the YOLO detection heads without any further dimensional adjustment.

This placement is significant because the final backbone stage is where contextual aggregation can affect all downstream detection scales. The design described for DRFSPPF therefore targets late-stage feature enrichment rather than early-stage local feature extraction. The paper frames this as an efficient way to improve multi-scale context modeling while preserving the lightweight character of the overall detector.

## 2. Internal architecture and processing stages

The DRFSPPF block proceeds through five explicit stages [2509.12442]. First, it performs multi-scale pooling and concatenation. The input feature map \(F\) is pooled with three square max-pool kernels \(k \in \{5,9,13\}\), each using stride \(=1\) and padding \(=(k-1)/2\) so that spatial size is preserved:
\[
P_k(F) = \mathrm{MaxPool}_{k \times k}(F).
\]
The original feature map and the three pooled outputs are concatenated along the channel dimension:
\[
F' = \mathrm{Concat}[F, P_5(F), P_9(F), P_{13}(F)] \in \mathbb{R}^{H \times W \times 4C}.
\]

Second, DRFSPPF applies a large-kernel directional depthwise convolution, denoted DConv. Two depthwise separable convolutions are applied in sequence: a vertical filter
\[
W^{(v)} \in \mathbb{R}^{(2d-1)\times 1 \times 4C}
\]
and a horizontal filter
\[
W^{(h)} \in \mathbb{R}^{1 \times (2d-1) \times 4C}.
\]
With dilation \(d=6\), \((2d-1)=11\), and the module computes
\[
F_{\mathrm{DConv}} = W^{(v)} * (W^{(h)} * F').
\]
The output remains \(H \times W \times 4C\). The stated purpose of this stage is to model long-range dependencies along each axis at very low cost.

Third, the block applies dilated depthwise convolution, denoted DDConv, again with channel-wise grouping:
\[
F_{\mathrm{DDConv}} = W^{(v)}_{\mathrm{dil}} * \left(W^{(h)}_{\mathrm{dil}} * F_{\mathrm{DConv}}\right),
\]
where
\[
W^{(v)}_{\mathrm{dil}} \in \mathbb{R}^{k' \times 1 \times 4C}, \qquad
W^{(h)}_{\mathrm{dil}} \in \mathbb{R}^{1 \times k' \times 4C},
\]
and the dilation rate is \(d\). In the paper, \(k' = \lfloor 11/d \rfloor\), typically \(1\) or \(2\), so that the dilated kernel still covers an \(11 \times 11\) effective window.

Fourth, DRFSPPF performs residual recalibration. A pointwise convolution
\[
W^{(1 \times 1)} \in \mathbb{R}^{1 \times 1 \times 4C \rightarrow C}
\]
projects \(F_{\mathrm{DDConv}}\) back to \(C\) channels, and the result is fused with the original feature map by element-wise multiplication:
\[
\bar{F} = (W^{1 \times 1} * F_{\mathrm{DDConv}}) \odot F.
\]
The description given in the paper is that this fuses local detail and global context.

Fifth, the output projection is produced by a final \(1 \times 1\) convolution:
\[
\mathrm{DRFSPPF}(F) = W^{\mathrm{out}} * \bar{F} \in \mathbb{R}^{H \times W \times C}.
\]
Because the only cross-channel mixing occurs through the \(1 \times 1\) convolutions, the block is explicitly characterized as lightweight.

## 3. Mathematical basis and receptive-field expansion

The paper formalizes the dilated convolution used in DRFSPPF for a single-channel input \(X \in \mathbb{R}^{H \times W}\), kernel \(K \in \mathbb{R}^{M \times N}\), and dilation rate \(d\) as
\[
Y = X \star_d K,
\]
with
\[
Y(i,j) = \sum_{m=1}^{M} \sum_{n=1}^{N} X(i + d \cdot m - o_m,\; j + d \cdot n - o_n) \cdot K(m,n),
\]
where \(o_m = \lfloor M/2 \rfloor\) and \(o_n = \lfloor N/2 \rfloor\). In multi-channel depthwise convolution, this sum is performed per channel [2509.12442].

The receptive-field analysis is stated explicitly. A standard convolution with kernel size \(k\) has
\[
RF_{\mathrm{standard}} = k.
\]
A \(k \times k\) convolution with dilation \(d\) expands its receptive field to
\[
RF_{\mathrm{dilated}} = k + (k-1)\cdot(d-1).
\]
For \(k=11\) and \(d=6\),
\[
RF_{\mathrm{dilated}} = 11 + 10 \cdot 5 = 61.
\]

The pooling stage is also formalized. For \(F \in \mathbb{R}^{H \times W \times C}\),
\[
T_1 = F, \qquad
T_2 = \mathrm{MaxPool}_{5 \times 5, s=1, p=2}(F), \qquad
T_3 = \mathrm{MaxPool}_{9 \times 9, s=1, p=4}(F), \qquad
T_4 = \mathrm{MaxPool}_{13 \times 13, s=1, p=6}(F),
\]
and the SPPF output becomes
\[
F' = \mathrm{Concat}[T_1, T_2, T_3, T_4] \in \mathbb{R}^{H \times W \times 4C}.
\]

On that basis, the paper contrasts DRFSPPF with standard SPPF. Standard SPPF pools capture receptive fields of up to \(13\), but rely on pure max-pooling. DRFSPPF adds a directional receptive field of \(11\) through large-kernel DConv, then applies a dilated convolution with \(k=11\) and \(d=6\) to yield a receptive field of \(61\). When cascaded, the overall block receptive field can exceed
\[
13 + (11-1) + (61-1) \approx 83
\]
pixels in each principal direction, which is described as a \(6\times\) increase over vanilla SPPF. The design rationale presented in the paper is that depthwise and directional separable convolutions make this receptive-field expansion possible at low cost because there is no additional channel mixing until the \(1 \times 1\) layers.

## 4. Computational cost and efficiency trade-offs

The computational-cost analysis is reported for \(640 \times 640\) input resolution [2509.12442]. The YOLOv11n baseline with vanilla SPPF requires \(6.4\) GFLOPs. The full Cott-ADNet, which includes DRFSPPF plus all other modules, requires \(7.5\) GFLOPs. A variant labeled "Remove DRFSPPF" requires \(7.3\) GFLOPs. The stated implication is that DRFSPPF alone adds approximately \(0.2\) GFLOPs over the version with everything else held constant.

The trade-off is also quantified in the source: \(+0.2\) GFLOPs, described as less than a \(3\%\) increase, is exchanged for a \(2.36\%\) absolute gain in mAP50 and \(2.34\%\) in \(F_1\). The latency increase is characterized as millisecond-level on a V100 and as not precluding real-time \(30+\) FPS deployment.

The efficiency argument depends on structural choices already built into the module. All convolutions in DRFSPPF are depthwise or \(1 \times 1\) pointwise, and the block preserves spatial dimensions, so no interpolation or reshaping is required. Within the terms used in the paper, the computational profile is therefore not based on reducing feature resolution, but on controlling arithmetic and memory overhead while enlarging the effective receptive field.

## 5. Empirical impact in ablation studies

The empirical effect of DRFSPPF is reported through an ablation comparing Cott-ADNet with and without the module [2509.12442]. Precision changes from \(90.486\%\) without DRFSPPF to \(91.543\%\) with DRFSPPF, a \(\Delta +1.057\%\). Recall changes from \(86.211\%\) to \(89.753\%\), a \(\Delta +3.542\%\). The \(F_1\)-score changes from \(88.297\%\) to \(90.639\%\), a \(\Delta +2.342\%\). mAP@0.5 changes from \(90.925\%\) to \(93.285\%\), a \(\Delta +2.360\%\). mAP \((0.5{:}0.95)\) changes from \(71.068\%\) to \(71.296\%\), a \(\Delta +0.228\%\).

These numbers locate the largest gains in recall, \(F_1\), and mAP@0.5, rather than in mAP \((0.5{:}0.95)\). This suggests that the module’s principal benefit, within the reported experiments, is not merely a marginal reshaping of precision-recall balance but a broader improvement in recovering target instances under the tested field conditions. The paper further states that the increases are well above typical random fluctuations and are consistent across multiple training runs, and that stable convergence curves confirm the statistical significance of the DRFSPPF gains.

The module’s contribution is situated within the broader detector results reported for Cott-ADNet as a whole: \(91.5\%\) Precision, \(89.8\%\) Recall, \(93.3\%\) mAP50, \(71.3\%\) mAP, and \(90.6\%\) \(F_1\)-Score with only \(7.5\) GFLOPs. The full system is also described as maintaining stable performance under multi-scale and rotational variations.

## 6. Integration, portability, and interpretive boundaries

The integration guidance given for DRFSPPF is explicit [2509.12442]. Any architecture that ends in an SPP or SPPF stage can swap in DRFSPPF. The only stated requirement is to match the input channel count \(C\) at insertion. The dilation rate \(d\) and the kernel sizes \(11\) and \(5/9/13\) may be tuned to reflect different input resolutions or object scales. Because the block’s only cross-channel mixing is via \(1 \times 1\) convolutions, it is said to slot cleanly between any feature extractor and detection head.

Deployment considerations are presented in similarly concrete terms. In practice, the entire Cott-ADNet, including DRFSPPF, runs at \(30+\) FPS on a single GPU at \(640 \times 640\), and can be quantized or converted to TensorRT for \(20+\) FPS on edge GPUs. The block preserves spatial dimensions, which removes the need for costly interpolation or reshaping, and this is part of the reason the module is framed as compatible with real-time use.

Several interpretive boundaries follow from the evidence given. A common misconception would be to treat DRFSPPF as a pure pooling variant. The architectural description shows that its behavior is not limited to pooled multi-scale aggregation; it also depends on directional depthwise convolution, dilated depthwise convolution, and residual recalibration. Another possible misconception is that receptive-field expansion in this setting necessarily requires expensive dense channel mixing. The source description instead emphasizes that no additional channel mixing occurs until the \(1 \times 1\) convolutions. A plausible implication is that the module’s utility derives from combining large effective receptive fields with a constrained channel-interaction budget rather than from any single operation in isolation.

Within the paper’s own summary, DRFSPPF extends vanilla SPPF by injecting large-kernel directional depthwise convolutions, following them with dilated depthwise convolutions for a receptive field exceeding \(80\) pixels, and fusing the result back to the original feature through lightweight residual recalibration. In the reported cotton-boll detection setting, that design yields a \(2.36\%\) boost in mAP@0.5 for a \(+0.2\) GFLOPs cost, while remaining suitable for real-time, resource-constrained deployment.

Source: https://www.emergentmind.com/topics/dilated-receptive-field-sppf