Improved Wavelet Scattering Deep Network

• The paper introduces a modified scattering transform that interleaves low- and high-frequency operations, resulting in superior feature extraction for UAV-based foreign object detection.
• It replaces the traditional CNN backbone with a fixed-wavelet scattering front-end and a compact MLP, significantly reducing computational cost while maintaining accuracy.
• The approach demonstrates robust performance with high classification accuracy (up to 95%) and real-time inference on edge devices for high-resolution UAV imagery.

An improved wavelet scatter deep network is a scattering-based deep architecture in which the classical cascade of wavelet filtering, modulus nonlinearity, and low-pass averaging is modified to improve discriminative power, robustness, or efficiency while retaining the analytic structure of scattering representations. In the specific formulation introduced for UAV inspection of power lines, the term denotes a lightweight classifier in which an Improved Wavelet Scattering Network (IWSN) replaces the convolution and pooling backbone of a CNN, and a three-layer fully connected network performs the final classification of foreign objects and background scenes (Zhichao et al., 15 Jul 2025).

1. Scattering-theoretic basis

Classical wavelet scattering networks arise from Mallat’s scattering transform, a deep architecture with fixed filters. For a 2D image $x$, the construction uses a scaling function $\phi$ and wavelet filters $\psi_\lambda$, then alternates convolution, modulus, and low-pass averaging. The zeroth-order coefficient is

$S_0 x(u) = x * \phi(u),$

the first-order coefficients are

$$U_1 x(u,\lambda_1) = |x * \psi_{\lambda_1}(t)|,\qquad S_1 x(u,\lambda_1) = U_1 x(u,\lambda_1) * \phi_1(u),$$

and higher orders iterate the same pattern,

$$U_n x(u,\lambda_1,\dots,\lambda_n) = \left|\, \big|x * \psi_{\lambda_1}\big| * \psi_{\lambda_2} \cdots * \psi_{\lambda_n}\right|, \qquad S_n x(u,\lambda_1,\dots,\lambda_n) = U_n x(u,\lambda_1,\dots,\lambda_n) * \phi_n(u).$$

This framework is valued for translation invariance induced by low-pass averaging, stability to small deformations, and energy-conservation-like behavior (Bruna et al., 2012).

Within this framework, scattering networks occupy a position between analytic signal processing and deep convolutional architectures. They resemble shallow CNNs in topology, but their filters are predefined rather than learned, and the nonlinearity is typically a complex or absolute-value modulus rather than ReLU. Subsequent developments have treated scattering as a platform for architectural modification rather than a fixed endpoint, producing a family of “improved” wavelet scattering networks that alter filter families, path structure, nonlinear regularization, or the learned head stacked above the scattering front-end (Oyallon et al., 2014).

2. Redefinition of high-order coefficients

The IWSN used for edge foreign-object recognition modifies the internal definition of high-order modulus coefficients. The paper states that “the positions of the scale function and the wavelet function are interchanged in the expression,” so the improved construction departs from the standard pattern of repeated high-pass filtering before final averaging. Its zeroth- and first-order terms remain

$$S_0 x = x * \phi(u), \qquad U_1 x(u,\lambda_1) = |x * \psi_{\lambda_1}(t)|, \qquad S_1 x(u,\lambda_1) = U_1 x(u,\lambda_1) * \phi_1(u),$$

but the second-order term is redefined as

$$U_2 x(u,\lambda_1,\lambda_2) = |x * \phi_{21}(u)| * \psi_{\lambda_2}(t), \qquad S_2 x(u,\lambda_1,\lambda_2) = U_2 x(u,\lambda_1,\lambda_2) * \phi_2(u),$$

and the general $m$-th-order term is written as

$$U_m x(u,\lambda_1,\dots,\lambda_m) = \Big|x * \phi(u) * \psi_{\lambda_1}(u) * \phi_{m-1}(u) * \psi_{\lambda_m}(t)\Big|,$$

$\phi$0

The intended effect is that the first $\phi$1 stages emphasize low-frequency or macroscopic structure, while the last step extracts high-frequency detail; this produces higher-order coefficients that combine global structure with local detail rather than only deep high-frequency chains (Zhichao et al., 15 Jul 2025).

For 2D images, the scaling and wavelet functions are built from separable filters,

$\phi$2

$\phi$3

with $\phi$4 the low-pass filter and $\phi$5 the high-pass filter. The implementation evaluates several biorthogonal wavelets—bior1.1, bior2.2, bior1.3, bior2.4, bior3.5, and bior2.6—and selects a small subset to balance accuracy and FLOPs. Biorthogonal wavelets are used because they allow linear phase, exact reconstruction with symmetric filters, and flexible trade-offs between smoothness and support. The design also decouples scattering depth from the number of distinct wavelet bases, allowing a three-layer scattering network to be built from a few repeated bases such as bior1.1, bior2.2, and bior1.3 (Zhichao et al., 15 Jul 2025).

The empirical rationale is task-specific. When object and background textures differ strongly, standard scattering coefficients can emphasize target detail; when target and background share similar fine texture or exhibit weak texture, the improved coefficients that retain low-frequency information can continue to highlight the foreign object. The paper reports that the improved network has higher overall accuracy and better precision/recall than the unimproved scattering network, with especially large gains for bird nests, which it identifies as the most dangerous class for faults (Zhichao et al., 15 Jul 2025).

3. Architecture and computational profile

The improved wavelet scatter deep network for transmission-line inspection is deliberately minimalist. The original images are RGB with resolutions up to $\phi$6 and sometimes $\phi$7, but only one color channel is used in the final model because the reported comparison of R, G, and B channels yields identical accuracy. The chosen input is the B channel, so the effective input is a single-channel 2D image (Zhichao et al., 15 Jul 2025).

The IWSN front-end replaces the convolution and pooling backbone of a CNN. It computes three scattering orders and produces modulus coefficients $\phi$8 together with scattering coefficients $\phi$9. The model primarily uses the three orders of modulus coefficients as classifier input. Spatial resolution is reduced by a factor of two at each scattering order in each dimension, so third-order features are described as being reduced to roughly $\psi_\lambda$0 of the original spatial area while preserving multi-scale content (Zhichao et al., 15 Jul 2025).

These coefficients are then flattened and passed into a very small multilayer perceptron. The paper states that three orders of modulus coefficients are flattened into a vector of length

$\psi_\lambda$1

The classifier has three fully connected layers with ReLU activations: $\psi_\lambda$2 with the dense-layer mapping

$\psi_\lambda$3

and ReLU

$\psi_\lambda$4

The model contains no convolutional layers, no pooling layers, no batch normalization, and no dropout after the scattering front-end. This compactness is central to its edge-computing orientation (Zhichao et al., 15 Jul 2025).

The reported computational cost is correspondingly small. Using bior1.1, IWSN feature extraction at $\psi_\lambda$5 requires approximately $\psi_\lambda$6M FLOPs, and the MLP requires approximately $\psi_\lambda$7 MFLOPs, giving roughly $\psi_\lambda$8 MFLOPs total. The paper contrasts this with a CNN example in which one convolutional layer alone costs approximately $\psi_\lambda$9–$S_0 x(u) = x * \phi(u),$0 GFLOPs and with YOLO-style detectors whose FLOPs exceed $S_0 x(u) = x * \phi(u),$1 GFLOPs at similar resolutions. This suggests that the architecture is intended less as a general detector than as a highly constrained recognition pipeline optimized for high-definition edge inference (Zhichao et al., 15 Jul 2025).

4. Foreign-object recognition and edge deployment

The immediate application is UAV patrol of high-voltage transmission lines, where foreign objects such as bird nests, kites, textiles, and plastic bags can induce short circuits, line breakage, or oscillation. The argument for edge deployment is operational rather than purely algorithmic: local inference avoids communication delay and loss, remains usable under harsh weather or network conditions, and supports adaptive flight and on-site decision making (Zhichao et al., 15 Jul 2025).

The dataset begins with 1697 transmission-line images containing foreign objects. Because the images are high resolution, additional background-only images are produced by cropping and by masking foreign objects with background textures. After augmentation by cropping, scaling, and rotating, the total training-plus-validation set contains 2764 images. The class proportions are reported as 63.5% bird nests, 10.8% kites, 14.6% textiles, and 11.1% plastics. For global evaluation, a separate 3000-image test set is constructed with 950 foreign-object images and 2050 background images (Zhichao et al., 15 Jul 2025).

On the 3000-image test set, the global foreign-object versus background task yields the following reported metrics:

Model	TPR / PPV / ACC	Note
SVM	80.4% / 93.6% / 92.0%	Handcrafted-feature baseline
YOLOv5s	81.7% / 98.5% / 93.8%	Highest precision
YOLOv8s	91.9% / 91.2% / 94.6%	Highest recall
IWSDN	88.9% / 94.8% / 94.9%	Highest overall accuracy

The reported interpretation is specific: YOLOv8s has the highest recall but more false positives, YOLOv5s has the highest precision but lower recall, and IWSDN attains the highest overall accuracy, $S_0 x(u) = x * \phi(u),$2 above YOLOv5s and $S_0 x(u) = x * \phi(u),$3 above YOLOv8s, while requiring much lower computational complexity (Zhichao et al., 15 Jul 2025).

Per-class results also remain above 90% across the four foreign-object types. Bird nests reach approximately 95% accuracy, kites approximately 94.7%, textiles achieve recall 96.3%, precision 97.5%, and accuracy 94.4%, and plastics achieve 94.4% accuracy. The comparison with the unimproved scattering model is also task-relevant: for bird nests, accuracy rises from 92.0% to 95.0%, and reported inference time at 720P decreases slightly from 6.7 ms to about 6.5 ms (Zhichao et al., 15 Jul 2025).

Deployment is reported on five edge devices: Orange Pi Zero3, Raspberry Pi 4B, Orange Pi 3B, Jetson Nano, and Orange Pi 5B. At $S_0 x(u) = x * \phi(u),$4, the reported frame rates are 37 FPS, 39.5 FPS, 36.4 FPS, 66.7 FPS, and 149.3 FPS, respectively, so the system exceeds 30 FPS on all tested devices. On Orange Pi 5B, the reported resolution-dependent behavior is 4.4 ms at $S_0 x(u) = x * \phi(u),$5, 6.7 ms at $S_0 x(u) = x * \phi(u),$6, and 14.9 ms at $S_0 x(u) = x * \phi(u),$7, corresponding to about 227 FPS, 149.3 FPS, and 67 FPS. The paper further notes that actual runtime is about ten times the theoretical estimate because of Python implementation and suboptimal system-level optimization; it explicitly suggests C/C++ or Rust ports and better hardware tuning as routes to lower latency (Zhichao et al., 15 Jul 2025).

5. Relation to other improved scattering architectures

The expression “improved wavelet scatter deep network” does not refer to a single canonical design. It names a family of attempts to preserve scattering’s invariance and stability while modifying the filter bank, path structure, or learned head. The UAV-oriented IWSDN improves scattering by redefining high-order modulus coefficients and by using a fixed-filter front-end plus a tiny MLP, but other lines of work alter different components (Zhichao et al., 15 Jul 2025).

One line replaces Mallat’s original wavelets with dual-tree complex wavelets. A dual-tree wavelet ScatterNet with a parametric log transformation uses DTCWT filters to improve computational efficiency and applies a scale-dependent log transform $S_0 x(u) = x * \phi(u),$8 to reduce skewness and aid feature selection; it is reported to outperform Mallat’s ScatterNet on image benchmarks in both classification accuracy and computational efficiency (Singh et al., 2017). A related multi-resolution DTCWT scattering network processes multiple input resolutions, uses regional $S_0 x(u) = x * \phi(u),$9 modulus nonlinearities, and outperforms Mallat’s ScatterNet on four datasets spanning images, audio, biology, and materials (Singh et al., 2017).

A second line increases filter diversity rather than altering coefficient definitions. The Sparse Multi-Family Deep Scattering Network crosses multiple wavelet families such as Morlet, Gammatone, and Paul across layers and adds a universal frame-thresholding operator that sparsifies latent coefficients and improves robustness to non-stationary noise. In bird-audio detection, the sparse multi-family variant improves mean AUC from 71.17 to 73.52 over the non-thresholded multi-family version (Cosentino et al., 2020).

A third line redesigns the underlying 2D analytic structure. The Monogenic Wavelet Scattering Network replaces Morlet filters with monogenic wavelets built from the Riesz transform, thereby encoding amplitude, phase, and orientation in a way intended to be more natural for 2D texture geometry. On CUReT, it reaches 97.34% accuracy, exceeding the best reported Morlet-based Kymatio STN configuration at 96.61% (Chak et al., 2022).

A fourth line introduces explicit parameter learning. Parametric scattering networks keep the Morlet form but learn scales, orientations, center frequencies, and aspect ratios through backpropagation. The reported result is that learned scattering parameterizations yield significant gains over standard fixed scattering in small-sample classification settings, while empirical deformation tests suggest that the stability of scattering is not substantially degraded (Gauthier et al., 2021).

These variants clarify that “improvement” in scattering networks can mean at least four distinct operations: changing the wavelet family, changing how paths are constructed or thresholded, learning wavelet geometry, or embedding scattering as a fixed front-end inside a larger learned architecture. The UAV-oriented IWSDN belongs to the fixed-front-end lineage, but its particular innovation is the low-/high-frequency interleaving of high-order coefficients rather than multi-family routing, monogenicity, or full parameter learning (Zhichao et al., 15 Jul 2025).

6. Limitations, misconceptions, and future directions

A common misconception is to treat all improved scattering networks as direct replacements for object detectors. The UAV model does not output bounding boxes or segmentation masks; it classifies whole images or crops into foreign-object categories or background. Operational deployment therefore presupposes either predefined crops or a separate localization mechanism. The paper explicitly identifies this distinction and suggests combining IWSN features with lightweight detection heads or using scattering features as a drop-in replacement for early convolutional layers in detectors (Zhichao et al., 15 Jul 2025).

Another misconception is that “improved” denotes a universally superior architecture. The comparative literature does not support that reading. Two-layer wavelet scattering can compete strongly with shallow pretrained CNN front-ends on object datasets, but deeper pretrained CNNs still achieve much higher accuracy on Caltech-101 and Caltech-256 than fixed two-layer scattering alone. This suggests that the label is architectural rather than absolute: it marks a refinement of scattering relative to a baseline, not a universal replacement for learned deep networks (Oyallon et al., 2013).

For the UAV-specific IWSDN, several limitations are explicit or strongly implied. The wavelet bases are manually chosen, the training set is small at roughly $$U_1 x(u,\lambda_1) = |x * \psi_{\lambda_1}(t)|,\qquad S_1 x(u,\lambda_1) = U_1 x(u,\lambda_1) * \phi_1(u),$$0–$$U_1 x(u,\lambda_1) = |x * \psi_{\lambda_1}(t)|,\qquad S_1 x(u,\lambda_1) = U_1 x(u,\lambda_1) * \phi_1(u),$$1k images, and runtime remains dominated by Python overhead rather than the theoretical FLOP count. The model is also tuned for recognition rather than dense localization, and while scattering is stable to small deformations, large viewpoint changes or heavy clutter are not directly studied. The paper’s future directions therefore emphasize optimized C/C++ or Rust implementations, hardware–software co-design for wavelet operations, larger and more varied datasets, extension to detection frameworks, and partial learning of wavelet parameters while preserving interpretability and stability (Zhichao et al., 15 Jul 2025).

More broadly, subsequent research suggests two complementary trajectories. One is to make scattering more expressive while staying fixed-filter—through monogenic, dual-tree, fractional, or multi-family constructions. The other is to keep the scattering topology but relax fixed filter geometry through learnable parameters or hybrid back-ends. A plausible implication is that the most durable role of improved wavelet scatter deep networks lies not in displacing end-to-end deep learning wholesale, but in supplying mathematically structured, low-compute, deformation-stable front-ends where data are limited, edge constraints are strict, or interpretability of multiscale features is operationally valuable (Gauthier et al., 2021).