IET Image Processing最新文献

Challenging underwater environmental conditions cause severe degradation of underwater images, including intensity attenuation and colour distortion, leading to incomplete feature representation and posing difficulties for state-of-the-art detectors. To address this issue, we propose a cross-channel feature fusion network (CCFF-Net), which innovatively enriches feature representation through complementary feature fusion across multiple channels. The network comprises three key components. First, an adaptive greyscale image generation module is designed to dynamically enhance channel information specialized for object appearance, selectively reinforcing detector-preferred cues in greyscale representation. Second, a cross-channel feature fusion module is introduced to facilitate information interaction between greyscale and chromatic channels, compensating for potential feature degradation caused by colour distortion and intensity attenuation by generating complementary and enhanced feature representations. Third, an enhanced feature pyramid network-based multi-scale feature fusion module is proposed to improve detection performance by reinforcing feature representations for both small-scale and occluded objects. Extensive experiments on four public datasets validate the effectiveness of CCFF-Net, achieving mAP improvements of 2.9%, 2.6%, 4.0% and 1.8% compared to the baseline on the DUO, UODD, UDD and URPC2020 datasets, respectively, demonstrating the superiority and generalization capability of the proposed method.

This study introduces a novel attention-guided Discrete Wavelet Transform (DWT)-based steganography framework, named Attention-Guided Feature Perturbation (AGFP), which integrates deep visual attention maps with transform-domain embedding to enhance imperceptibility, robustness, and steganalysis resistance. Unlike recent deep-learning-based steganographic systems such as iSCMIS, JARS-Net, and RMSteg, which achieve high visual fidelity but are susceptible to statistical detection, AGFP perturbs only those wavelet coefficients that are identified as perceptually and statistically stable by attention mechanisms extracted from pre-trained CNN models (VGG19, ResNet50, AlexNet, and GoogLeNet). The proposed method is evaluated on the USC-SIPI dataset and the BOSSBase 1.01 benchmark. Experimental results show that AGFP achieves PSNR values between 64.29 and 55.43 dB and SSIM scores between 0.9999 and 0.9989 across varying payloads, indicating consistently high visual quality. While iSCMIS reports slightly higher PSNR and SSIM values, AGFP significantly outperforms all compared methods in bit error rate (BER)—achieving 0.01–0.12, compared to 0.45–0.47 for iSCMIS, 0.31–0.37 for RMSteg, and 0.57–0.75 for JARS-Net. Furthermore, AGFP attains the lowest RS, SPA, and SRM steganalysis detection scores among both classical and deep-learning-based systems. These results confirm that AGFP offers a more balanced and secure steganographic solution, combining high imperceptibility with substantially enhanced robustness and detectability resistance, positioning it as a strong alternative to recent deep-learning-based steganographic frameworks.

The analysis of posterior scleral images is crucial for understanding the progression of myopia and developing effective interventions. However, due to the subtle edge variations in posterior scleral images and their high memory requirements, existing 3D models face limitations in memory consumption, while 2D models struggle to capture cross-slice spatial information, making it difficult to balance the performance and memory usage. Therefore, we propose SFM-UNet, a novel model designed specifically for OCT analysis of the posterior sclera. SFM-UNet integrates frequency and spatial information through a cross-slice spatial-frequency memory module, effectively capturing the complex spatial relationships in 3D images. We conducted experiments on a private dataset EyePS2024 and the publicly available BraTS2020 dataset, where SFM-UNet demonstrated competitive performance across all datasets, demonstrating its effectiveness and practicality in posterior sclera OCT analysis.

In urban road scenes, the cross-domain data distribution differences caused by light and weather changes make the generalization performance of single-domain trained object detectors in unknown weather scenarios significantly degraded (e.g., daytime-sunny trained models in dusk-rainy scenarios reduce the detection accuracy by more than 46% on average); moreover, faster R-CNN suffers from insufficient generalization capability and inefficient real-time inference due to architectural constraints. To address the above challenges, this paper proposes the CEN-RTDETR method to improve the single-domain generalization capability through a collaborative enhancement strategy: CP-Mix color channel permutation dynamically simulates the RGB color channel permutation to simulate the multi-sky color bias phenomenon, and enhances the color robustness of the input data; NP feature normalization perturbation applies a random feature perturbation to the channel statistics of the shallow feature map to optimize the extraction of texture, color and other basic style features; CORAL Loss minimizes the difference in feature distribution between the source domain and the virtual target domain through second-order statistical matching. The experimental results show that CEN-RTDETR achieves significant performance improvement on the cross-weather scenario dataset DWD: The mean average precision ([email protected]) across different weather scenarios increases from 39.66% to 42.72% (+3.06%), especially for dusk-rainy and night-rainy scenarios, where [email protected] rises from 32.9% and 18.6% to 39.0% (+6.1%) and 24.1% (+5.5%) in extreme weather. The method in this paper effectively solves the cross-domain generalization and efficiency problems in single-domain generalized object detection, which provides new technical possibilities for real-time detection in complex urban road scenes.

The assessment of segmentation quality plays a fundamental role in the development, optimization, and comparison of segmentation methods which are used in a wide range of applications. With few exceptions, quality assessment is performed using traditional metrics, which are based on counting the number of erroneous pixels but do not capture the spatial distribution of errors. Established distance-based metrics such as the directed average Hausdorff distance ($�\mathrm{dAHD}$) are difficult to interpret and compare for different methods and datasets. In this paper, we introduce the surface consistency coefficient ($�\mathrm{SCC}$), a novel distance-based quality metric that quantifies the spatial distribution of errors based on their proximity to the surface of the structure. Through a rigorous analysis using synthetic data and real segmentation results, we demonstrate the robustness and effectiveness of $�\mathrm{SCC}$ in distinguishing errors near the surface from those farther away. At the same time, $�\mathrm{SCC}$ is easy to interpret and comparable across different structural contexts.

Multi-focus image fusion (MFIF) aims to synthesise an all-in-focus image from multiple source images captured with different focal depths. However, effectively distinguishing focused regions from defocused backgrounds remains challenging due to the smooth transitions and lack of ground-truth supervision. MFIF synthesises an all-in-focus image by selecting sharp regions from images captured at different focal planes. However, many deep methods ignore the gradient discrepancy between focused and defocused areas during multi-scale extraction, treat features without quantifying their reliability, and build decision maps using only horizontal and vertical gradients, leading to detail loss and structural misclassification. We propose MSFE-Net, a gradient-aware multi-scale feature enhancement network comprising three core components. First, a gradient‑aware multi‑scale feature enhancement module (GradientAwareMFE) balances fine detail and contextual cues by fusing dual atrous spatial pyramid pooling branches with small and large dilation rates under a learned gradient gate aligned to the Sobel gradient magnitude. Second, focus-reliability attention converts local-variance statistics into a spatial reliability mask and applies squeeze-and-excitation to modulate channels. Third, enhanced spatial frequency fusion integrates four-directional gradients with local context and morphological refinement to produce a robust decision map for pixel-wise fusion. Trained in a self-supervised manner on Microsoft Common Objects in Context (MS-COCO) using mean squared error, structural similarity, and gradient-consistency losses, and evaluated on Lytro, MFFW, and MFI-WHU datasets, MSFE-Net attains state-of-the-art or competitive results on most metrics, delivering sharper edges and fewer artefacts.

With the rapid development of image editing technology, image manipulation localisation is facing increasingly severe challenges. Traditional methods cannot effectively capture tampering features at different scales and fail to fully utilise the tampering trace information at the boundary of the manipulated regions, resulting in limited localisation accuracy. To address these issues, we propose MEN-IML, a multi-scale edge-aware network for image manipulation localisation. First, we design an efficient multi-scale edge-aware feature fusion module. This module enhances the expressive capability of edge details by assigning learnable weights to features at different scales and incorporating an edge feature awareness mechanism. Second, we adopt depthwise separable convolutional decoders instead of the traditional multi-layer perceptron decoders, which not only focus on extracting spatial features of each channel, but also improve the model's ability to integrate cross-channel information. Finally, we design a multi-scale edge loss to supervise the boundaries of manipulated regions at multiple scales, effectively enhancing the sensitivity of the model to manipulated region boundaries. Cross-dataset experimental results on multiple public datasets demonstrate that, compared to mainstream MVSS-Net++ and the state-of-the-art IML-ViT, the proposed method improves average performance by 13.5% and 6.3%, effectively enhancing the localisation accuracy of image tampering.

In pharmaceutical manufacturing, tablets experience flaws like half-grain tablets, multi-pill tablets and paste tablets as a result of breaking, adhesion or inadequate pressing. The packaging material, typically composed of dual aluminium foil, exhibits high reflectivity. This reduces the overall image contrast and weakens the feature differences between defective areas and the background. This phenomenon significantly affects the accuracy of defect detection in pharmaceutical quality control. To address the above problems, this paper proposes a contrast enhancement method called Wavelet Quadtree Contrast Limited Adaptive Histogram Equalisation Tablet Enhancement (WCTE) for tablet images. The method combines Haar Wavelet Transform and CLAHE with adaptive quadtree chunking to enhance the detection accuracy of low-contrast tablets. The proposed approach uses Haar wavelets to decompose tablet images at multiple scales. It then applies power transform enhancement and soft-threshold denoising to sub-bands of different frequencies, highlighting edge details and suppressing background interference. The image is reconstructed by the inverse wavelet transform. To avoid edge artefacts and local over-enhancement from fixed blocks, an adaptive quadtree CLAHE strategy driven by local variance is applied. This allows adaptive segmentation of enhancement regions and ensures smooth transitions. The YOLOv11 model is utilised to identify the target in the augmented tablet image, facilitating the precise classification of typical flaws such as half-grain tablets, multi-pill tablets and paste tablets. Experimental results show that WCTE-enhanced images outperform both traditional CLAHE and unenhanced ones in entropy, contrast, clarity and average gradient. The comprehensive scores improve by 27.15% and 11.42% relative to the original and CLAHE images, respectively. The YOLOv11 model demonstrates a significant enhancement in defect detection accuracy for WCTE-enhanced images, with improvements of 13.33% and 9.06% for half-grain tablets and paste-like defects, respectively. The accuracy for multi-pill defects remains stable, while the overall mean average precision (mAP) increases by 2.5%, and the false detection rate decreases by 9%. This improvement further substantiates the efficacy and applicability of this method in low-contrast target detection tasks.

Recent trends in computational pathology favour increasingly complex deep learning architectures, raising the question of whether such complexity is necessary for routine diagnostic tasks. This study challenges this assumption through a comprehensive analysis of the relationship between model complexity, data pre-processing, and performance across four fundamental digital pathology tasks: nuclei counting, steatosis quantification, glomeruli detection, and Ki67 proliferation index (PI) assessment. We evaluated five deep learning models of varying complexity (lightweight: MobileNetV2, U-Net, and more complex: ConvNeXt, K-Net, and Swin Transformer) combined with different image pre-processing techniques. To evaluate model performance without extensive ground truth (GT) annotations, we introduced a validation strategy utilizing the relative absolute deviation (RAD) between network predictions and correlation of performance metrics. Our findings demonstrate that pre-processing strategies, particularly stain normalization (NORM), can be more impactful than model complexity, reducing error rates by up to 50% compared to processing original (ORIG) images. With appropriate pre-processing, lightweight models achieved comparable or superior results to complex models while reducing processing times by up to 40%. Only specific tasks involving complex morphological features, such as glomeruli detection, significantly benefited from more sophisticated architectures. This study provides an evidence-based framework for selecting optimal model-pre-processing combinations in clinical settings, suggesting that investing in pre-processing pipelines rather than model complexity may be more beneficial for routine computational pathology applications.

Face recognition is a cornerstone of smart system development and remains a challenging research domain, particularly in achieving accurate and real-time identification of facial expression. Existing approaches often struggle to manage dynamic variations in pose, illumination and expression, resulting in reduced accuracy, delayed processing and inconsistent identification. The research difficulties are addressed by introducing a facial recognition framework based on the unified optimised feature vector (UOFV) to enhance robustness and efficiency in constrained environments. The framework combines standard descriptors (e.g., LBP, HOG and HDG) with deep learning features (e.g., CNN-based embeddings), uniting local texture, directional intensity patterns and high-level semantic representations into a single discriminative feature space. These extracted features are then optimised using the binary grey wolf optimisation algorithm, selected for its strong balance between exploration and exploitation. The optimisation systematically selects the most relevant attributes while discarding redundancies, reducing dimensionality without compromising discriminative power and producing the optimised UOFV. The framework is implemented in MATLAB and validated on six public datasets: ORL, YALE, warpPIE10P, dbCMUfaces, LFW and CASIA-WebFace. For classification, six standard machine learning models (e.g., KNN, DT, SVM, NB, DA and RF) were used to identify facial images. Experimental results confirm the effectiveness of the UOFV approach, achieving 99.12% accuracy on ORL, 87.87% on YALE, 100% on warpPIE10P, 99.91% on dbCMUfaces, 98.56% on LFW and 98.48% on CASIA-WebFace. Moreover, the approach demonstrates remarkable efficiency, with an average execution time of only 1.2 milliseconds per recognition task, confirming its suitability for real-time applications. A demonstration code of the proposed framework is publicly available at: https://www.mathworks.com/matlabcentral/fileexchange/178149-face-recognition-using-optimized-feature-extraction-and-ml