Blueprint Separable Convolutions

• Blueprint separable convolutions are defined as a reparameterization of standard convolution kernels, using a 1×1 pointwise convolution followed by a depthwise convolution with a shared spatial blueprint to drastically reduce parameters.
• They exploit intra-kernel correlations, with PCA analyses showing that up to 50% of variance is captured by a single component, which validates the blueprint factorization approach.
• Empirical studies demonstrate that BSConv improves performance in tasks like image classification, super-resolution, and infant cry detection while reducing parameters compared to standard and depthwise separable convolutions.

Blueprint separable convolutions (BSConv) are efficient convolutional operators that re-parameterize regular convolutional kernels to exploit intra-kernel correlations, replacing a dense $M \times K \times K$ kernel bank by channel-mixing weights and a 2D blueprint per output channel. In operational form, BSConv is implemented as a $1\times1$ pointwise convolution followed by a depthwise $K\times K$ convolution; in kernel form, a filter is written as $\mathbf{K}_n[m,:,:] = w_{n,m}\cdot \mathbf{B}_n$ or, equivalently in the infant-cry formulation, $F_c^{(n)}=\omega_c^{(n)}\cdot B^{(n)}$. The operator was introduced as a highly efficient building block for CNNs and has subsequently been used as a drop-in replacement for standard convolutions in MobileNet-like classifiers, efficient image super-resolution networks, and noisy infant cry detection systems (Haase et al., 2020, Li et al., 2022, Yu et al., 26 Aug 2025).

1. Formal definition and parameterization

A standard convolutional layer maps an input tensor of size $M \times Y \times X$ to an output tensor of size $N \times Y \times X$ using $N$ kernels of size $M \times K \times K$. Its parameter count is therefore

$M \cdot N \cdot K^2.$

BSConv assumes that, for each output filter, the slices along the input-channel axis are scalar-modulated versions of a common spatial template. In the formulation emphasized in the core BSConv paper, this is written as

$1\times1$0

where $1\times1$1 is a $1\times1$2 blueprint for filter $1\times1$3 and $1\times1$4 is a scalar weight for input channel $1\times1$5 (Haase et al., 2020).

The infant-cry detection paper expresses the same idea as

$1\times1$6

with $1\times1$7 denoting a pointwise convolution weight and $1\times1$8 a 2D blueprint used in a depthwise convolution. Under this factorization, the parameter count becomes

$1\times1$9

in contrast to the standard-convolution count $K\times K$0 (Yu et al., 26 Aug 2025).

The implementation follows directly from the factorization. First, a $K\times K$1 pointwise convolution linearly combines the input channels; second, a depthwise $K\times K$2 convolution applies one spatial blueprint per output channel. The BSRN paper describes this explicitly as

$K\times K$3

and characterizes BSConv as “essentially the inverse of Depthwise Separable Convolution (DSConv),” since DSConv applies depthwise spatial filtering first and pointwise channel mixing second (Li et al., 2022).

A concrete scale comparison appears in the infant-cry study for $K\times K$4 and $K\times K$5: a standard convolution requires $K\times K$6 parameters, whereas BSConv requires $K\times K$7 parameters (Yu et al., 26 Aug 2025).

2. Statistical motivation: intra-kernel correlations

The principal theoretical motivation for BSConv is the empirical dominance of intra-kernel correlations in trained convolutional filters. Quantitative analyses of kernels from trained models such as VGG-19 and ResNet-50 showed that slices of a filter along the depth axis are often highly similar and differ mainly by scalar factors, sometimes with inverted sign. The same study reports PCA analyses in which the variance explained by the first principal component is often about $K\times K$8, supporting the use of a single spatial blueprint distributed along depth by channel-dependent weights (Haase et al., 2020).

This motivation distinguishes BSConv from the interpretation commonly applied to depthwise separable convolutions. The BSConv paper argues that DSC-based architectures such as MobileNets implicitly rely on cross-kernel correlations, whereas the BSConv formulation is based on intra-kernel correlations and therefore allows a more efficient separation of regular convolutions (Haase et al., 2020). In that sense, BSConv is not merely a computational rearrangement of an existing operator; it is a model of kernel redundancy.

A common simplification is to describe BSConv as only “pointwise-then-depthwise.” That description is operationally correct, but incomplete. The stronger claim made in the literature is that the order of operations is justified by the statistical structure of learned kernels: channel mixing first produces intermediate channels that can each be convolved with a blueprint, because much of the original $K\times K$9 kernel bank is redundant under an intra-kernel-correlation model (Haase et al., 2020).

The same idea reappears in application papers in a more task-specific form. In the infant-cry detector, BSConv is described as exploiting intra-kernel correlations to reduce both parameter count and computational cost over traditional and depthwise separable convolutions, while remaining efficient and expressive for structured inputs such as log Mel-spectrograms (Yu et al., 26 Aug 2025).

3. Variants and architectural integration

BSConv has been used both as a standalone convolutional primitive and as the basis for further compressed variants. The core paper introduces a low-rank “subspace BSConv” in which the pointwise weight matrix $\mathbf{K}_n[m,:,:] = w_{n,m}\cdot \mathbf{B}_n$0 is factorized as

$\mathbf{K}_n[m,:,:] = w_{n,m}\cdot \mathbf{B}_n$1

with $\mathbf{K}_n[m,:,:] = w_{n,m}\cdot \mathbf{B}_n$2, $\mathbf{K}_n[m,:,:] = w_{n,m}\cdot \mathbf{B}_n$3, and $\mathbf{K}_n[m,:,:] = w_{n,m}\cdot \mathbf{B}_n$4. To reduce redundancy in the learned basis, an orthonormal regularization term is added: $\mathbf{K}_n[m,:,:] = w_{n,m}\cdot \mathbf{B}_n$5 and the total loss is written as $\mathbf{K}_n[m,:,:] = w_{n,m}\cdot \mathbf{B}_n$6 (Haase et al., 2020).

In image super-resolution, the BSRN work uses the BSConvU variant as a drop-in replacement for standard convolution inside residual blocks. The operator forms the core of the Blueprint Shallow Residual Block and is also used in the initial feature extraction and enhanced spatial attention modules. In that context, BSConv is presented as an efficient two-layer stack—$\mathbf{K}_n[m,:,:] = w_{n,m}\cdot \mathbf{B}_n$7 convolution followed by depthwise $\mathbf{K}_n[m,:,:] = w_{n,m}\cdot \mathbf{B}_n$8 convolution—that reduces redundancy while preserving the ability to exploit channel correlations (Li et al., 2022).

In noisy infant cry detection, BSConv replaces standard convolutions in the encoder and in the attention modules. The reported architecture is a multi-scale convolutional recurrent neural network operating on log Mel-spectrograms and augmented with an efficient spatial attention mechanism and a contrast-aware channel attention module. Multiple BSConv blocks with pooling are used for feature extraction, features from different encoder blocks are concatenated to preserve spatial and frequency details, and the resulting representation is passed to a time-frequency recurrent neural network for adaptive denoising before classification by an MLP (Yu et al., 26 Aug 2025).

The integration pattern is therefore broader than simple backbone replacement. BSConv appears in residual feature extractors, attention modules, multi-scale encoders, and low-rank channel-mixing pipelines, which suggests that the blueprint factorization is compatible with several common CNN design motifs.

4. Empirical performance across application domains

In large-scale and fine-grained image classification, BSConv was reported to “clearly and consistently improve MobileNets and other DSC-based architectures without introducing any further complexity.” On ImageNet, MobileNetV1(1.0) improved from $\mathbf{K}_n[m,:,:] = w_{n,m}\cdot \mathbf{B}_n$9 to $F_c^{(n)}=\omega_c^{(n)}\cdot B^{(n)}$0, and on CIFAR-10 MobileNetV1(1.0) improved from $F_c^{(n)}=\omega_c^{(n)}\cdot B^{(n)}$1 to $F_c^{(n)}=\omega_c^{(n)}\cdot B^{(n)}$2. The same study reports improvements of up to $F_c^{(n)}=\omega_c^{(n)}\cdot B^{(n)}$3 percentage points on fine-grained datasets and states that, when used as a drop-in replacement for standard architectures such as ResNets, BSConv variants outperform their vanilla counterparts by up to $F_c^{(n)}=\omega_c^{(n)}\cdot B^{(n)}$4 percentage points on ImageNet (Haase et al., 2020).

In efficient image super-resolution, BSConvU was compared with both standard convolution and depthwise separable convolution inside an RFDN-style baseline. The reported numbers were: RFDN, $F_c^{(n)}=\omega_c^{(n)}\cdot B^{(n)}$5K parameters and $F_c^{(n)}=\omega_c^{(n)}\cdot B^{(n)}$6G Multi-Adds; RFDN-DSConv, $F_c^{(n)}=\omega_c^{(n)}\cdot B^{(n)}$7K parameters and $F_c^{(n)}=\omega_c^{(n)}\cdot B^{(n)}$8G; RFDN-BSConvU, $F_c^{(n)}=\omega_c^{(n)}\cdot B^{(n)}$9K parameters and $M \times Y \times X$0G. On Set5, the corresponding PSNR / SSIM values were $M \times Y \times X$1 for RFDN, $M \times Y \times X$2 for RFDN-DSConv, and $M \times Y \times X$3 for RFDN-BSConvU. The BSRN model itself reported $M \times Y \times X$4dB PSNR / $M \times Y \times X$5 SSIM with $M \times Y \times X$6K parameters / $M \times Y \times X$7G MACs on Set5 $M \times Y \times X$8 SR, compared with RFDN’s $M \times Y \times X$9dB / $N \times Y \times X$0 and $N \times Y \times X$1K parameters / $N \times Y \times X$2G MACs; the smaller BSRN-S variant won first place in the model complexity track of the NTIRE 2022 Efficient SR Challenge (Li et al., 2022).

In noisy infant cry detection, the BSConv-based model was reported to have the lowest parameter count and the second lowest FLOPs among the compared methods: $N \times Y \times X$3 million parameters and $N \times Y \times X$4 giga FLOPs, versus MobileNetV2’s $N \times Y \times X$5 million and $N \times Y \times X$6 giga, ECAPA-TDNN’s $N \times Y \times X$7 million and $N \times Y \times X$8 giga, BiLSTM’s $N \times Y \times X$9 million and $N$0 giga, and ResNet18’s $N$1 million and $N$2 giga. Detection performance under varying SNR was reported as follows for the BSConv model: accuracy $N$3, $N$4, $N$5, and $N$6 for clean, $N$7 dB, $N$8 dB, and $N$9 dB; F1-score $M \times K \times K$0, $M \times K \times K$1, $M \times K \times K$2, and $M \times K \times K$3 for the same conditions. The next-best ResNet18 F1-scores were $M \times K \times K$4, $M \times K \times K$5, $M \times K \times K$6, and $M \times K \times K$7, while BiLSTM reported $M \times K \times K$8, $M \times K \times K$9, $M \cdot N \cdot K^2.$0, and $M \cdot N \cdot K^2.$1. The study further states that the F1 drop from clean to $M \cdot N \cdot K^2.$2 dB was $M \cdot N \cdot K^2.$3 for the BSConv model, compared with $M \cdot N \cdot K^2.$4 for ResNet18, and that removing the BSConv-enhanced module, the multi-scale connections, or the RNN denoising module all reduced performance (Yu et al., 26 Aug 2025).

Taken together, these results position BSConv as an efficiency-oriented convolutional primitive whose gains are not limited to a single benchmark family. The reported advantages span ImageNet classification, fine-grained recognition, super-resolution, and low-SNR audio detection.

5. Relation to depthwise separable and other separable operators

The closest baseline for BSConv is depthwise separable convolution. DSConv applies depthwise spatial filtering per channel and then performs pointwise channel mixing; BSConv reverses that order, performing channel mixing first and then spatial filtering. The BSRN paper treats this inversion as a key architectural difference and ties it to better utilization of channel correlations (Li et al., 2022).

The broader conceptual background is provided by work on separable convolutions generally. “Towards a New Interpretation of Separable Convolutions” does not directly discuss BSConv, but it reinterprets standard depthwise separable convolutions as a hybrid of ResNeXt-like grouped spatial filtering and Inception-like $M \cdot N \cdot K^2.$5 cross-channel analysis, rather than as an “extreme Inception” limit (Ghosh, 2017). This provides context for BSConv because BSConv changes the decomposition itself: it does not retain the standard depthwise-then-pointwise module, but instead substitutes a blueprint-based factorization motivated by intra-kernel correlations.

The literature also distinguishes BSConv from other efficient operators that separate computation along different axes. The sequence-to-sequence speech-recognition paper on time-depth separable convolutions states explicitly that TDS is not a blueprint separable convolution: TDS separates temporal aggregation and depth/frequency mixing at the block level, whereas blueprint separable convolution is a kernel-factorization approach (Hannun et al., 2019). That distinction matters because “separable” in modern CNN design can refer to different objects: channelwise spatial filtering, low-rank kernel structure, temporal-depth decoupling, or hardware-aware one-dimensional decompositions.

A recurrent misconception is therefore terminological. BSConv belongs to the separable-convolution family only in a broad taxonomic sense. Its defining feature is not separability per se, but blueprint factorization based on intra-kernel correlations.

6. Design implications, limitations, and adjacent directions

The empirical record in the cited work supports three recurring design implications. First, BSConv is routinely used as a drop-in replacement for standard convolution or DSC components, which lowers parameter count and computational cost while preserving or improving accuracy in the reported tasks (Haase et al., 2020, Li et al., 2022, Yu et al., 26 Aug 2025). Second, BSConv is compatible with attention mechanisms, residual learning, and multi-scale aggregation, as shown in BSRN and the infant-cry detector (Li et al., 2022, Yu et al., 26 Aug 2025). Third, the low-rank subspace variant indicates that the pointwise stage itself may contain additional compressible structure beyond the primary blueprint factorization (Haase et al., 2020).

The published results also indicate limitations. The BSRN paper describes BSConv as “slightly less flexible than standard conv,” even while arguing that it is less redundant and makes better use of channel correlations (Li et al., 2022). More generally, the BSConv formulation presupposes that a shared blueprint plus channelwise scaling is an adequate approximation to a dense kernel. Where that assumption is weak, a plausible implication is that the operator may lose some of the representational freedom of standard convolution.

Adjacent work suggests two further directions. “Separable Layers Enable Structured Efficient Linear Substitutions” shows that pointwise convolutions are fully connected linear layers and can be replaced by structured decompositions such as ACDC, Tensor-Train, Tucker, low-rank, HashedNet, and Linearised ShuffleNet; this suggests a possible further compression path for BSConv architectures, whose first stage is itself a pointwise convolution (Gray et al., 2019). In a different line, FuSeConv was proposed as a hardware-aware drop-in replacement for depth-wise separable convolution because DWSConv is inefficient on systolic arrays; this indicates that the most effective separable operator can depend on the target accelerator, not only on statistical efficiency (Selvam et al., 2021).

Within the current literature, BSConv is best understood as a statistically motivated kernel factorization that occupies a specific position in the broader ecology of efficient convolutions: more structured than standard convolution, differently structured from DSConv, and empirically effective across classification, restoration, and spectrogram-based detection tasks (Haase et al., 2020, Li et al., 2022, Yu et al., 26 Aug 2025).