End-to-End Image Compression with Segmentation Guided Dual Coding for Wind Turbines

Raül Pérez-Gonzalo; Andreas Espersen; Søren Forchhammer; Antonio Agudo

cs.CVMar 31, 2026

End-to-End Image Compression with Segmentation Guided Dual Coding for Wind Turbines

Authors: Raül Pérez-Gonzalo, Andreas Espersen, Søren Forchhammer, Antonio Agudo

Organizations: Wind Power LAB, 1150 Copenhagen, Denmark · Institut de Robòtica i Informàtica Industrial, CSIC-UPC, 08028 Barcelona, Spain · Department of Photonics Engineering, Technical University of Denmark (DTU), 2800 Lyngby, Denmark

Abstract

Transferring large volumes of high-resolution images during wind turbine inspections introduces a bottleneck in assessing and detecting severe defects. Efficient coding must preserve high fidelity in blade regions while aggressively compressing the background. In this work, we propose an end-to-end deep learning framework that jointly performs segmentation and dual-mode (lossy and lossless) compression. The segmentation module accurately identifies the blade region, after which our region-of-interest (ROI) compressor encodes it at superior quality compared to the rest of the image. Unlike conventional ROI schemes that merely allocate more bits to salient areas, our framework integrates: (i) a robust segmentation network (BU-Netv2+P) with a CRF-regularized loss for precise blade localization, (ii) a hyperprior-based autoencoder optimized for lossy compression, and (iii) an extended bits-back coder with hierarchical models for fully lossless blade reconstruction. Furthermore, our ROI framework removes the sequential dependency in bits-back coding by reusing background-coded bits, enabling parallelized and efficient dual-mode compression. To the best of our knowledge, this is the first fully integrated learning-based ROI codec combining segmentation, lossy, and lossless compression, ensuring that subsequent defect detection is not compromised. Experiments on a large-scale wind turbine dataset demonstrate superior compression performance and efficiency, offering a practical solution for automated inspections.

Figures & tables

Fig. 1 : Overview of our ROI algorithm for blade image compression . Our method begins by extracting a segmentation mask using our proposed BU-Netv2+P. This mask is converted into a polygonal representation by classifying image patches as either blade (ROI) or background. The polygon mask is then efficiently encoded via its corner points. The background is highly compressed with our lossy coder, while the blade region is coded either lossily –in better quality and less distortion– or losslessly. In ROI lossless case, the background is first encoded to provide its lossy bitstream for parallel bits-back coding.

Fig. 2 : BU-Netv2+P algorithm pipeline to segment blades : 1) an enhanced BU-Netv2, 2) a hole filling step, 3) a random forest module that exploits the input image set and the preliminary solution and, 4) a repeated hole filling operation.

Fig. 3 : Hole filling scheme. Example of artifact removal inside the segmented blade region ( $\mathbf{\hat{s}}^{BU}$ ). The method detects blade orientation to locate border edges (in red) and applies a standard hole-filling algorithm [ 19 ] to produce the mask $\mathbf{\hat{s}}^{H}$ .

Panel of Figure 4 — Fig. 4 : Overview of HP+EASN-deep hierarchical model and architecture. Left-side illustrates the 2-level nested variable model, where latent variables $\mathbf{z}_{1}$ and $\mathbf{z}_{2}$ are obtained during encoding $h_{\boldsymbol{\phi}}$ and quantized $Q$ . Right-side shows the architecture, a symmetric 2-layer autoencoder composed of convolutional and up/downsampling blocks with non-linear activations.

Fig. 5 : Network architecture corresponding to the decoder transforms : $p_{\boldsymbol{\theta}}(\mathbf{x}|\mathbf{z}_{1})$ and $p_{\boldsymbol{\theta}}(\mathbf{z}_{l}|\mathbf{z}_{l+1})$ for $l\in\{1,\ldots,L-1\}$ . The arrows indicate the direction of the forward pass.

Fig. 6 : Order of the encoding and decoding operations for a standard hierarchical bits-back and for its optimized scheme on a $L$ -level nested variable model for lossless compression.

Panel of Figure 7 — Fig. 7 : Acceptance-ratio curves of each segmentation step. $\Theta^{accuracy}$ and $\Theta^{recall}$ similarity metrics are displayed in the top and bottom sides, respectively. On the right column, the lefts plots are zoomed on the black dashed area.

Performance	BU-Netv2	Fill holes	Random forest	Fill holes 2
metric	${\hat{\mathbf{s}}}^{BU}$	$\hat{\mathbf{s}}^{H1}$	$\hat{\mathbf{s}}^{RF}$	$\hat{\mathbf{s}}\equiv\hat{\mathbf{s}}^{H2}$
Accuracy [%]	96.82	96.98	97.54	97.61
Recall [%]	96.41	96.61	97.54	97.67
F1-score [%]	95.60	95.72	96.44	96.50

Table I : Performance results of each segmentation step .

Method	Accuracy	Recall	F1-score	mIoU	Relative	Relative
	[%]	[%]	[%]	[%]	accuracy	recall
U-Net [ 29 ]	86.24	68.93	77.95	75.94	1	1
DeepLabv3+ [ 27 ]	94.14	87.38	89.03	87.47	1.09	1.27
SW [ 28 ]	93.48	91.71	91.37	87.44	1.08	1.33
ResNeSt [ 34 ]	94.23	91.47	92.77	89.63	1.09	1.33
U-NetFormer [ 35 ]	96.20	93.51	94.42	91.75	1.12	1.36

Table II : Quantitative comparison on blade segmentation with respect to state-of-the-art methods. Models marked with an asterisk only propose segmentation regions without a class; results assume a perfect region classifier.

Panel of Figure 10 — Fig. 8 : Acceptance-ratio curves of comparison techniques. $\Theta^{accuracy}$ and $\Theta^{recall}$ similarity metrics are displayed in the left and right parts, respectively.

Panel of Figure 11 — Fig. 9 : Rate-distortion curves for state-of-the-art lossy coders . Learning-based curves for each distortion loss metric: PSNR (left) or MS-SSIM (right) with red regions zoomed (middle). Non-learning-based coders are depicted with dashed lines.

	Performance		Compression Time				Decompression Time
Coding	BD-rate	BD-rate	Minimum	Maximum	Mean	Slope	Minimum	Maximum	Mean	Slope
scheme	[PSNR]	[MS-SSIM]	time [s]	time [s]	time [s]	growth [s/bit]	time [s]	time [s]	time [s]	growth [s/bit]
HP [ 11 ]	15.35	5.81	5.76	6.17	5.96	0.25	6.51	15.60	12.83	5.57
HP+EASN-deep [ 21 ]	0	0	6.46	6.95	6.64	0.26	7.86	15.35	12.74	4.00
JA [ 51 ]	2.55	-6.04	200.68	204.19	203.01	1.96	442.74	458.26	451.08	8.69
JA+EASN-deep [ 21 ]	2.14	-1.11	200.53	205.75	203.42	2.83	443.54	458.08	453.10	7.88

Table III : BD-rate, compression and decompression times of a full-resolution image for state-of-the-art lossy coders.

Coding	Mean	Std
scheme	bit rate [bit/pix]	bit rate [bit/pix]
Bit-swap	8.98	1.70
PNG [ 42 ]	11.33	1.64
BPG [ 45 ]	11.61	2.19
JPEG2000 [ 10 ]	10.41	1.69
WebP [ 43 ]	10.17	1.77

Table IV : Compression performance of a full-resolution image for the distinct state-of-the-art lossless coders. Std denotes the standard deviation across test images.

	Compression		Decompression
Coding	Mean	Std	Mean	Std
scheme	time [s]	time [s]	time [s]	time [s]
Bit-swap	5688	21.56	4367	19.64
PNG [ 42 ]	11.35	2.28	9.8	1.96
BPG [ 45 ]	7.22	0.49	14.03	1.99
JPEG2000 [ 10 ]	7.03	0.89	18.78	4.91

Table V : Compression and decompression times of a full-resolution image for state-of-the-art lossless coders. Std denotes the standard deviation across test images.

Panel of Figure 15 — Fig. 10 : Proportion of ROI width along the pressure side of a wind turbine blade . Red numbers denote the $i$ -th image.

Panel of Figure 16 — Fig. 11 : Blade pictures compressed with our proposed ROI algorithm . The first column showcases the images compressed using the lossy HP+EASN-deep coder with two distinct rate-distortion models on the blade ( $[\zeta=0.1]\cdot$ MSE) and background ( $[\zeta=0.01]\cdot$ MSE), whereas the third column presents these blade regions compressed losslessly. Second and fourth column represent their left column images augmented and gray-scaled, so the high-frequency components can be visualized and compared. The sub-captions indicate the PSNR, MS-SSIM and bit/px. See Fig. 8 from Supplementary for the original images.

Fig. 1 : Four example instances of the blade imagery dataset . The complexity on this data arises due to the background variation and wild illumination conditions.

Fig. 2 : Validation accuracy after applying the random forest step with distinct sets of input pixels . The performance is analyzed in terms of the number of local pixel neighbors taken as input $n$ and the distance $d$ between these input pixels. The purple horizontal line represents the accuracy obtained before applying the random forest step.

Fig. 3 : Windfarm test performance of each segmentation step . For readability, the plot does not include every windfarm.

Panel of Figure 20 — Fig. 4 : Qualitative comparison evaluation between BU-Netv1+P and our approach BU-Netv2+P . The figure depicts on the left side the test images $\mathbf{x}$ and its ground truth mask $\mathbf{s}$ ; on the inner side the segmentation steps of BU-Netv1+P (from left to right columns: BU-Netv1, first hole filling, random forest, second hole filling); and on the right side the the steps of our proposed segmentation algorithm BU-Netv2+P (from left to right columns: BU-Netv2 ${\hat{\mathbf{s}}}^{BU}$ , first hole filling $\hat{\mathbf{s}}^{H1}$ , random forest ensemble $\hat{\mathbf{s}}^{RF}$ and second hole filling $\hat{\mathbf{s}}\equiv\hat{\mathbf{s}}^{H2}$ ).

Panel of Figure 21 — Fig. 5 : Failure cases on test images after running each segmentation step . First column: input color image $\mathbf{x}$ . From second to fifth column: BU-Netv2 ${\hat{\mathbf{s}}}^{BU}$ , first hole filling $\hat{\mathbf{s}}^{H1}$ , random forest ensemble $\hat{\mathbf{s}}^{RF}$ and second hole filling $\hat{\mathbf{s}}\equiv\hat{\mathbf{s}}^{H2}$ estimations, respectively. On both sides, the same information is displayed.

Fig. 6 : Computational time in seconds of an individual image for each segmentation step . Computational time for the BU-Net, hole filling and random forest steps are reported. The blue line represents the cumulative time.

Fig. 7 : Additional bits per pixel of information required to decompress each rate-distortion model of HP+EASN-deep.

Panel of Figure 25 — Fig. 8 : Quality assessment of two blade pictures with high compression rates . The first row showcases the complete original images, whereas the second row presents a zoomed-in image accompanied by two different compression rates. Sub-captions indicate the distortion loss term and the compression performance. Red regions indicated with numbers pinpoint the following artifacts: 1) blocking artifacts, 2) smoothed contours, 3) smoothed textures and 4) visual color scale change.

Panel of Figure 26 — Fig. 9 : Compression and decompression time of a full-resolution image for the distinct state-of-the-art lossy coders.

Panel of Figure 27 — Fig. 10 : Cumulative average bit rate of the distinct coding steps when applying the recursive bits-back algorithm (left) or the optimized version (middle). Average net bit rate of the distinct coding steps (right). The minimum negative bit rate denotes the required initial bits for a single image patch.

Fig. 11 : Bit rate as a function of the number of image patches that have been sequentially compressed.

Panel of Figure 29 — Fig. 12 : Learning curves of each individual distribution when training on the validation dataset. Notice the inference curves are minimized when our desire is to maximize them. Top: Inference distributions (decoding steps). Bottom: Likelihood and prior distributions (encoding steps).

Fig. 13 : Ablation study of the number of latent variables $L$ and the patch size $PS$ over the Bit-swap model.

Fig. 14 : Compression and decompression times in terms of the number of latent variables $L$ and the patch size $PS$ over the Bit-swap model.

Panel of Figure 32 — Fig. 15 : Rate-distortion curves per region of ROI-eML ; HP+EASN-deep trained for squared error. The error bars indicate the std distortion on the compression performance. PSNR (left) and MS-SSIM (right) quality measures.

Coding	Region	Bit rate	Bit rate
scheme		mean	std
JPEG2000-ROI [ 10 ]	Blade	10.06	1.68
JPEG2000-ROI [ 10 ]	Background	10.31	1.72
ROI-eML	Blade	8.84	1.68
ROI-eML	Background	9.06	1.72

Table I : Lossless compression performance comparison of ROI-eML per region , implemented by means of Bit-swap.

Learned Image Compression

Learned image compression uses neural encoders, quantization, and entropy models to reduce image bitrates while balancing fidelity, perceptual quality, and task utility.

See all 120 papers

BladeYOLO: Wind Turbine Blade Defect Detection with Limited Annotations and Weak-Saliency Awareness
Jul 30, 2026Yabin Xu, Fangtao Zhang, Fan Wang +5Object Detection Structural Health Monitoring
School of Mechanical Engineering, Zhejiang Sci-Tech University, Hangzhou 310018, China · Department of Artificial Intelligence and Robotics, Zhejiang Energy Digital Technology Co., Ltd, Hangzhou 311100, China · School of Computer Science and Technology, Nanjing University of Aeronautics and Astronautics, Nanjing, China +1
ARCHE: Autoregressive Residual Compression with Hyperprior and Excitation
Mar 10, 2026Sofia Iliopoulou, Dimitris Ampeliotis, Athanassios SkodrasLearned Image Compression Autoregressive Modeling
Dept. of Electrical and Computer Engineering, University of Patras, Patras, Greece · Dept. of Digital Media and Communication, Ionian University, Argostoli, Greece
SLIM: Semantic-based Low-bitrate Image compression for Machines
Dec 20, 2025Hyeonjin Lee, Jun-Hyuk Kim, Jong-Seok LeeLearned Image Compression Generative Image Compression
School of Integrated Technology, Yonsei University, Republic of Korea · Department of Artificial Intelligence, Chung-Ang University, Republic of Korea

Jul 30, 2026J/K move · Enter open · S save

BladeYOLO: Wind Turbine Blade Defect Detection with Limited Annotations and Weak-Saliency Awareness

Yabin Xu, Fangtao Zhang, Fan Wang, Zhan Wang+4

School of Mechanical Engineering, Zhejiang Sci-Tech University, Hangzhou 310018, China · Department of Artificial Intelligence and Robotics, Zhejiang Energy Digital Technology Co., Ltd, Hangzhou 311100, China · School of Computer Science and Technology, Nanjing University of Aeronautics and Astronautics, Nanjing, China+1

Object Detection Structural Health Monitoring

Open paper AskPDF arXiv

Wind turbine blade defect detection remains highly challenging in real-world inspection scenarios due to limited on-site data and the subtle visual characteristics of defects. In practice, blade defects are often small-scale, low-contrast, and difficult to distinguish from complex backgrounds, which significantly limits the robustness of existing detectors. To address these challenges, we propose BladeYOLO, a defect detection framework for wind turbine blades. Specifically, we integrate a Vision Transformer (ViT) backbone initialized with DINOv3 self-supervised pre-trained weights into YOLOv12-L, enabling the transfer of large-scale generic visual priors to blade defect detection and improving feature representation under limited training annotations. To enhance the perception of subtle defects, we further develop a Mamba-guided Weak-Defect Enhancement module, which consists of a Detail-Enhanced Multi-scale Branch for preserving high-frequency structural cues and a Cross-Mamba module for progressively propagating high-level semantic guidance to shallow features. In addition, we introduce a lightweight Style-Injector module that captures environment-related style information via Fourier decomposition and injects it into selected ViT self-attention layers, thereby improving robustness against environment-induced appearance variations. Extensive experiments demonstrate that BladeYOLO achieves superior performance on the WTBlade-Defect dataset, with additional annotation-budget experiments showing its favorable performance under reduced training annotations. Evaluation on the public Wind Surface Defect dataset further provides supportive evidence for the cross-dataset robustness of BladeYOLO. In particular, on this public dataset, BladeYOLO outperforms the best competing method by 3.5% in mAP

_{50}

and 2.5% in mAP

_{50-95}

Abstract

Learned Image Compression

Explore similar work

BladeYOLO: Wind Turbine Blade Defect Detection with Limited Annotations and Weak-Saliency Awareness

ARCHE: Autoregressive Residual Compression with Hyperprior and Excitation

SLIM: Semantic-based Low-bitrate Image compression for Machines