Inverted Residual Depthwise Convolution Block

• The paper introduces IRDCB in HierLight-YOLO by replacing the YOLOv8 C2f module, achieving 22.1% fewer parameters and 9.4% reduction in FLOPs with negligible accuracy loss.
• IRDCB is a lightweight convolutional building unit that combines an inverted residual structure with two stacked depthwise convolutions and a conditional residual connection.
• Its efficiency arises from separating channel mixing using 1x1 convolutions for expansion and compression from spatial filtering handled by depthwise convolutions.

The Inverted Residual Depthwise Convolution Block (IRDCB) is a lightweight convolutional building unit that combines an inverted residual structure, depthwise-separable computation, and a conditional residual connection. In the specific usage documented for HierLight-YOLO, IRDCB is introduced as a replacement for the original YOLOv8-style C2f module in order to reduce parameters and computational cost while preserving detection accuracy in UAV imagery dominated by small objects (Chen et al., 26 Sep 2025). More broadly, IRDCB belongs to the same architectural family as the MobileNetV2 inverted residual block, whose defining pattern is expansion in a higher-dimensional latent space, depthwise spatial filtering, and projection back to a thin bottleneck representation (Sandler et al., 2018). Related works also explore alternative orderings and reallocations of the same basic ingredients, including the Depthwise-Pointwise-Depthwise (DPD) inverted bottleneck block and the asymmetrical bottleneck (Li et al., 2019, Yang et al., 2021).

1. Terminological scope and architectural lineage

The term IRDCB is used explicitly in HierLight-YOLO, where it denotes one of the two “innovative lightweight modules” used to improve efficiency in a YOLOv8-based detector for UAV photography (Chen et al., 26 Sep 2025). In that paper, IRDCB is presented as a redesign of the backbone feature-extraction block, intended to reduce model size and FLOPs without sacrificing the fine-grained spatial information needed for small-object detection.

At the level of architectural lineage, IRDCB is directly connected to MobileNetV2, whose central building unit is described as an inverted residual block with linear bottleneck (Sandler et al., 2018). MobileNetV2 inverts the classical bottleneck pattern: instead of connecting wide tensors and compressing internally, it starts from a thin bottleneck tensor, expands it, applies the nonlinear transform in the expanded space, then projects back to a thin output, with shortcut connections between the thin bottlenecks.

The broader family of related blocks includes variants that preserve the inverted-bottleneck idea while modifying internal operator order or the allocation of computation. The DPD block replaces pointwise expansion with depthwise convolution with channel multiplier $m>1$, yielding the sequence DWC $(m>1)$ $\to$ PWC $\to$ DWC $(m=1)$ (Li et al., 2019). AsymmNet keeps the standard pointwise–depthwise–pointwise macro-structure but changes the balance between the two pointwise layers through feature reuse and an asymmetry rate $r$ (Yang et al., 2021).

This suggests that “IRDCB” is best understood as a specific named instance within a larger class of inverted residual / depthwise-convolution bottleneck blocks, rather than as a universally standardized term. In the available sources, the exact name is tied to HierLight-YOLO, whereas MobileNetV2 supplies the canonical template and later works supply closely related variants.

2. Canonical block structure

In MobileNetV2, the practical inverted residual block has the tensor transformation

$$h \times w \times k \xrightarrow{1\times 1} h \times w \times (tk) \xrightarrow{3\times 3,\ s=s} \frac{h}{s} \times \frac{w}{s} \times (tk) \xrightarrow{\text{linear }1\times 1} \frac{h}{s} \times \frac{w}{s} \times k'$$

where the first $1\times1$ convolution expands channels by factor $t$, the middle $3\times3$ operator is depthwise, and the final $(m>1)$0 projection is linear (Sandler et al., 2018). The default expansion factor reported for MobileNetV2 is $(m>1)$1.

In HierLight-YOLO, the IRDCB is also a three-stage transform, but its middle stage is deeper. The block takes an input tensor $(m>1)$2 with input channels $(m>1)$3 and output channels $(m>1)$4, and is described as

$(m>1)$5

with the intended operational sequence stated as:

1. Expand with a $(m>1)$6 convolution
2. Filter with two stacked $(m>1)$7 depthwise convolutions
3. Compress with a final $(m>1)$8 convolution
4. Add a residual shortcut if input and output channels match (Chen et al., 26 Sep 2025)

The expansion stage increases channels from $(m>1)$9 to

$\to$0

The two depthwise convolutions each use a $\to$1 kernel, stride 1, padding 1, and group number $\to$2, so each channel is filtered independently. The compression stage maps $\to$3 back to $\to$4 using another $\to$5 convolution (Chen et al., 26 Sep 2025).

Relative to MobileNetV2, the distinguishing change is therefore not the presence of the inverted residual principle itself, but the use of two consecutive depthwise convolutions instead of one. The paper states that this improves nonlinear representation and local pattern capture (Chen et al., 26 Sep 2025).

3. Computational rationale and efficiency model

The efficiency logic underlying IRDCB-style blocks comes from the separation of channel mixing and spatial filtering. In a standard convolution block, a regular $\to$6 convolution performs both simultaneously, so each output channel interacts with all input channels. By contrast, inverted residual depthwise blocks use $\to$7 pointwise convolutions for channel expansion and compression, while using depthwise $\to$8 convolutions for spatial filtering (Sandler et al., 2018, Chen et al., 26 Sep 2025).

For MobileNetV2, the paper gives the cost of depthwise separable convolution as

$\to$9

compared with standard convolution cost

$\to$0

and states that for $\to$1, depthwise separable convolutions are roughly 8–9× cheaper than full convolutions, with only a small accuracy drop (Sandler et al., 2018).

For a MobileNetV2 bottleneck block with input size $\to$2, expansion factor $\to$3, kernel size $\to$4, input channels $\to$5, and output channels $\to$6, the multiply-add cost is

$\to$7

which formalizes why the expensive channel-mixing operations are confined to $\to$8 layers while the spatial operator remains depthwise (Sandler et al., 2018).

In HierLight-YOLO, the IRDCB efficiency argument is stated more directly through complexity comparisons. The expansion $\to$9 convolution has cost $(m=1)$0, whereas a standard convolution at the same channel dimensions would cost $(m=1)$1. Each depthwise convolution has complexity

$(m=1)$2

whereas a standard convolution over the expanded channels would cost

$(m=1)$3

This is identified as the core source of savings (Chen et al., 26 Sep 2025).

A closely related efficiency argument appears in the DPD block literature. For a feature map $(m=1)$4 expanded to $(m=1)$5, a PWC needs $(m=1)$6 parameters and computational cost $(m=1)$7, whereas a DWC with kernel size $(m=1)$8 and output channels $(m=1)$9 needs $r$0 parameters and cost $r$1. The parameter and computation ratios are both

$r$2

and for $r$3, the argument is that depthwise convolution is much more efficient than pointwise convolution for channel expansion when $r$4 is much larger than $r$5 (Li et al., 2019).

Taken together, these formulations support a consistent interpretation: the practical advantage of IRDCB-style blocks comes from delegating spatial modeling to depthwise operators and restricting dense cross-channel mixing to $r$6 layers, thereby reducing parameters, FLOPs, and activation cost.

4. Residual topology, bottlenecks, and information flow

A defining property of inverted residual blocks is that the shortcut operates between thin bottleneck tensors, not between wide expanded tensors. MobileNetV2 states that “the shortcut connections are between the thin bottleneck layers” and that shortcuts connecting bottlenecks perform better than shortcuts connecting expanded layers (Sandler et al., 2018). Residual addition is used only when the block preserves both spatial resolution and channel count; in practice, if stride $r$7 and input/output channels match, then

$r$8

otherwise there is no residual addition (Sandler et al., 2018).

The IRDCB in HierLight-YOLO follows the same conditional logic, formulated explicitly as

$r$9

Thus, the block behaves as a residual unit only when channel dimensions are compatible (Chen et al., 26 Sep 2025).

MobileNetV2 also links this residual topology to the concept of linear bottlenecks. The paper argues that applying nonlinearity in a narrow space can collapse information lying on a low-dimensional manifold, and therefore “it is important to remove non-linearities in the narrow layers in order to maintain representational power” (Sandler et al., 2018). Its core design principle is therefore nonlinearity in the wide layer and no nonlinearity in the narrow layer.

No explicit activation function or normalization layer is specified in the IRDCB subsection of HierLight-YOLO. The paper notes that such layers may exist implicitly in a YOLOv8 framework, but they are not explicitly defined in the IRDCB description (Chen et al., 26 Sep 2025). Accordingly, the available evidence supports a structural characterization of IRDCB, but not a more detailed claim about its activation or normalization policy.

A common misconception is to treat all inverted residual blocks as interchangeable. The sources do not support that view. MobileNetV2’s canonical block, HierLight-YOLO’s IRDCB, the DPD block, and AsymmNet’s asymmetrical bottleneck all share the inverted-residual/depthwise design space, but they differ materially in the placement of depthwise operators, the allocation of channel expansion, and the precise role of residual and projection layers (Sandler et al., 2018, Li et al., 2019, Yang et al., 2021, Chen et al., 26 Sep 2025).

5. Major variants and closely related blocks

Several architectures modify the inverted residual depthwise block while preserving its lightweight design objective.

Block	Core operator order	Distinctive feature
MobileNetV2 inverted residual	$$h \times w \times k \xrightarrow{1\times 1} h \times w \times (tk) \xrightarrow{3\times 3,\ s=s} \frac{h}{s} \times \frac{w}{s} \times (tk) \xrightarrow{\text{linear }1\times 1} \frac{h}{s} \times \frac{w}{s} \times k'$$0 expand $$h \times w \times k \xrightarrow{1\times 1} h \times w \times (tk) \xrightarrow{3\times 3,\ s=s} \frac{h}{s} \times \frac{w}{s} \times (tk) \xrightarrow{\text{linear }1\times 1} \frac{h}{s} \times \frac{w}{s} \times k'$$1 $$h \times w \times k \xrightarrow{1\times 1} h \times w \times (tk) \xrightarrow{3\times 3,\ s=s} \frac{h}{s} \times \frac{w}{s} \times (tk) \xrightarrow{\text{linear }1\times 1} \frac{h}{s} \times \frac{w}{s} \times k'$$2 DW $$h \times w \times k \xrightarrow{1\times 1} h \times w \times (tk) \xrightarrow{3\times 3,\ s=s} \frac{h}{s} \times \frac{w}{s} \times (tk) \xrightarrow{\text{linear }1\times 1} \frac{h}{s} \times \frac{w}{s} \times k'$$3 linear $$h \times w \times k \xrightarrow{1\times 1} h \times w \times (tk) \xrightarrow{3\times 3,\ s=s} \frac{h}{s} \times \frac{w}{s} \times (tk) \xrightarrow{\text{linear }1\times 1} \frac{h}{s} \times \frac{w}{s} \times k'$$4 projection	Linear bottleneck and shortcut between thin layers
DPD block	DWC $$h \times w \times k \xrightarrow{1\times 1} h \times w \times (tk) \xrightarrow{3\times 3,\ s=s} \frac{h}{s} \times \frac{w}{s} \times (tk) \xrightarrow{\text{linear }1\times 1} \frac{h}{s} \times \frac{w}{s} \times k'$$5 $$h \times w \times k \xrightarrow{1\times 1} h \times w \times (tk) \xrightarrow{3\times 3,\ s=s} \frac{h}{s} \times \frac{w}{s} \times (tk) \xrightarrow{\text{linear }1\times 1} \frac{h}{s} \times \frac{w}{s} \times k'$$6 PWC $$h \times w \times k \xrightarrow{1\times 1} h \times w \times (tk) \xrightarrow{3\times 3,\ s=s} \frac{h}{s} \times \frac{w}{s} \times (tk) \xrightarrow{\text{linear }1\times 1} \frac{h}{s} \times \frac{w}{s} \times k'$$7 DWC $$h \times w \times k \xrightarrow{1\times 1} h \times w \times (tk) \xrightarrow{3\times 3,\ s=s} \frac{h}{s} \times \frac{w}{s} \times (tk) \xrightarrow{\text{linear }1\times 1} \frac{h}{s} \times \frac{w}{s} \times k'$$8	Uses depthwise convolution for channel expansion
Asymmetrical bottleneck	PW1 adjustment $$h \times w \times k \xrightarrow{1\times 1} h \times w \times (tk) \xrightarrow{3\times 3,\ s=s} \frac{h}{s} \times \frac{w}{s} \times (tk) \xrightarrow{\text{linear }1\times 1} \frac{h}{s} \times \frac{w}{s} \times k'$$9 feature reuse/Concat $1\times1$0 DW $1\times1$1 PW2	Reallocates computation from PW1 to PW2
IRDCB in HierLight-YOLO	$1\times1$2 expand $1\times1$3 DW $1\times1$4 DW $1\times1$5 $1\times1$6 compress	Uses two consecutive depthwise convolutions

The DPD block is the closest case in which depthwise convolution is promoted from a middle spatial operator to the expansion mechanism itself. The first DWC uses channel multiplier $1\times1$7 to increase channels and complete downsampling, the PWC combines information of different channels and decreases the number of channels, and the final DWC filters each channel (Li et al., 2019). The paper concludes that “extracting spatial features is more important than combining channel information” and reports that networks with more DWC layers outperform networks with more $1\times1$8 convolution layers.

AsymmNet instead revisits the two pointwise convolutions in a standard MobileNetV2-style block. It proposes to reduce the first pointwise convolution, enrich information flow by direct feature reuse, and migrate the saved computations to the second pointwise convolution (Yang et al., 2021). The resulting block keeps the familiar inverted-residual macro-structure but changes the internal balance of capacity and channel mixing. For stride $1\times1$9, the theoretical cost ratio is given as

$t$0

which simplifies to

$t$1

The stated implication is that computation can be reallocated while total cost remains approximately unchanged (Yang et al., 2021).

These variants clarify that the design space around IRDCB is not limited to a single operator sequence. Different works target different bottlenecks: MobileNetV2 emphasizes information preservation in narrow layers, DPD emphasizes replacing pointwise expansion with depthwise expansion, AsymmNet emphasizes asymmetrical allocation between the two pointwise layers, and HierLight-YOLO emphasizes stronger local pattern capture through two stacked depthwise convolutions.

6. Empirical performance and documented trade-offs

The most direct empirical evidence for IRDCB comes from HierLight-YOLO. In the module comparison reported in Table 4, the results are:

Module	Accuracy / AP	Model size / cost
C3	44.1 AP$t$2	9.8M params, 36.0 GFLOPs
C2f	44.9 AP$t$3	11.3M params, 38.1 GFLOPs
IRDCB	44.8 AP$t$4	8.8M params, 34.5 GFLOPs

The paper summarizes this as 22.1% fewer parameters than the original C2f module, 9.4% fewer FLOPs than C2f, and no meaningful accuracy loss (Chen et al., 26 Sep 2025). Relative to C2f, the AP$t$5 difference is only 0.1%, while parameters fall from 11.3M to 8.8M and FLOPs from 38.1G to 34.5G.

The full ablation chain reports that baseline YOLOv8s achieves 43.0 AP$t$6, 26.0 AP$t$7, 11.1M params, and 28.5 GFLOPs, while the version with P2 head + HEPAN + IRDCB reaches 47.1 AP$t$8, 29.0 AP$t$9, 8.8M params, and 34.5 GFLOPs (Chen et al., 26 Sep 2025). The paper specifically notes that IRDCB reduces model size while maintaining comparable accuracy.

The IRDCB expansion study identifies $3\times3$0 as the best setting, with 44.9 AP$3\times3$1, 8.8M params, and 34.5 GFLOPs. Increasing $3\times3$2 from 2 to 4 with fixed $3\times3$3 yields only a small gain from 44.0 to 44.6 AP$3\times3$4, while parameters rise from 8.8M to 9.4M; for $3\times3$5, increasing $3\times3$6 beyond 2 degrades performance from 44.9 at $3\times3$7 to 44.5 at $3\times3$8 and 44.4 at $3\times3$9 (Chen et al., 26 Sep 2025). The paper concludes that too much expansion hurts optimization and efficiency.

The broader literature supports the general utility of this architectural family. MobileNetV2 reports a main model with 3.4M parameters and 300M multiply-adds, and states that it achieves better accuracy than MobileNetV1 at comparable or lower cost (Sandler et al., 2018). DPDNet reports that the number of parameters is only about 60% of that of MobileNetV2 for networks with the same number of layers, while achieving approximated accuracy (Li et al., 2019). In one representative setting, DPDNet uses 0.09M parameters and 12.6M FLOPs versus 0.14M parameters and 23.3M FLOPs for MobileNetV2, with 90.42% on CIFAR-10 and 66.36% on CIFAR-100 compared with MobileNetV2’s 89.92% and 67.07% (Li et al., 2019).

These results do not imply that all IRDCB-like blocks dominate all alternatives on all tasks. Rather, they document a recurring trade-off: lightweight inverted residual depthwise blocks can reduce parameters and FLOPs substantially while maintaining similar accuracy, but the optimal internal expansion and operator ordering remain architecture- and task-dependent.

7. Interpretation, significance, and recurrent points of debate

The principal significance of IRDCB lies in its role as a lightweight feature-extraction block that seeks to preserve representational adequacy under tight compute and memory budgets. In the detection setting of HierLight-YOLO, the block is motivated by the need to remain efficient on resource-constrained platforms without losing the spatial detail required for UAV small-object detection (Chen et al., 26 Sep 2025). In the mobile-classification setting of MobileNetV2, the same family is motivated by the need to separate input/output bottleneck dimension from the expressiveness of the nonlinear transformation (Sandler et al., 2018).

One recurrent point of debate concerns the relative importance of spatial extraction versus channel mixing. The DPD paper explicitly argues that “extracting spatial features is more important than combining channel information” and uses the performance of depthwise-heavy variants to support that interpretation (Li et al., 2019). By contrast, AsymmNet argues that the second pointwise convolution is especially important for learning cross-channel feature correlations, and therefore computation should be migrated toward PW2 rather than left in PW1 (Yang et al., 2021). These positions are not identical; they reflect different emphases within the same low-cost design space.

Another frequent misunderstanding is to equate “inverted residual” with a single immutable recipe. The available evidence shows at least four distinct realizations: the canonical MobileNetV2 block with one depthwise layer and linear projection; the IRDCB of HierLight-YOLO with two depthwise layers; the DPD block with depthwise expansion; and the AsymmNet block with asymmetrical pointwise allocation (Sandler et al., 2018, Chen et al., 26 Sep 2025, Li et al., 2019, Yang et al., 2021). Their commonality lies in the use of thin bottlenecks, lightweight spatial filtering, and conditional shortcutting, but their internal mechanics differ enough to affect both efficiency and accuracy.

A plausible implication is that IRDCB is best treated not as a singular fixed artifact, but as a design point within a family of inverted residual depthwise bottlenecks. In the sources considered here, its most concrete and fully named instantiation is the YOLOv8-backbone module in HierLight-YOLO, while its conceptual foundations are anchored in MobileNetV2 and its subsequent variants.