Image Quality Assessment (IQA) with references plays an important role in optimizing and evaluating computer vision tasks. Traditional methods assume that all pixels of the reference and test images are fully aligned. Such Aligned-Reference IQA (AR-IQA) approaches fail to address many real-world problems with various geometric deformations between the two images. Although significant effort has been made to attack Geometrically-Disparate-Reference IQA (GDR-IQA) problem, it has been addressed in a task-dependent fashion, for example, by dedicated designs for image super-resolution and retargeting, or by assuming the geometric distortions to be small that can be countered by translation-robust filters or by explicit image registrations. Here we rethink this problem and propose a unified, non-training-based Deep Structural Similarity (DeepSSIM) approach to address the above problems in a single framework, which assesses structural similarity of deep features in a simple but efficient way and uses an attention calibration strategy to alleviate attention deviation. The proposed method, without application-specific design, achieves state-of-the-art performance on AR-IQA datasets and meanwhile shows strong robustness to various GDR-IQA test cases. Interestingly, our test also shows the effectiveness of DeepSSIM as an optimization tool for training image super-resolution, enhancement and restoration, implying an even wider generalizability.

Low-level image processing has long been evaluated mainly from the perspective of visual fidelity. However, with the rise of deep learning and generative models, processed images may preserve perceptual quality while altering semantic content, making conventional Image Quality Assessment (IQA) insufficient for semantic-level assessment. In this paper, we formalize Semantic Similarity as a new evaluation task for low-level image processing, aimed at measuring whether semantic content is preserved after processing. We further present a structured formulation of image semantics based on semantic entities and their relations, and discuss the desired properties and constraints of a valid semantic similarity index. Based on this formulation, we propose Triplet-based Semantic Similarity Score (T3S), which models image semantics through foreground entities, background entities, and relations. T3S combines semantic entity extraction, foreground-background disentanglement, and open-world class/relation modeling. Experiments on COCO and SPA-Data show that T3S consistently outperforms existing fidelity-oriented metrics and representative semantic-level baselines, while better reflecting progressive semantic changes under diverse degradations. These results highlight the importance of semantic assessment in modern low-level vision.

In underwater environments, computer vision and image processing techniques have attracted increasing attention due to advances in marine biology and underwater vehicles. As a fundamental task, underwater species recognition plays a crucial role in various real-world applications. With the rapid development of deep learning, significant progress has been achieved in this field. However, existing surveys lack a systematic analysis of data-related challenges and are often limited to small-scale datasets. In this paper, we present a comprehensive survey of recent deep learning-based methods for underwater species recognition. We organize existing approaches into three main categories and provide a structured review of representative methods. Furthermore, we highlight fine-grained recognition and long-tailed distributions as key yet under-explored challenges, and analyze their impact through empirical evaluation. We also introduce large-scale benchmark datasets and provide unified baseline comparisons. Finally, we discuss emerging research directions and open problems to guide future studies.

Transmitting images through narrow-bandwidth underwater acoustic channels presents a significant challenge. However, for many applications, transmitting the entire image is not necessary. Instead, conveying only the critical information is typically sufficient to ensure task success. To address this, we have introduced a preprocessing strategy at the image acquisition stage, which involves down-sampling or feature extraction to distill the data into its essential components. This strategy is guided by utility considerations to ensure that the transmitted data is both concise and informative. However, existing quality evaluation methods are inadequate for providing accurate utility assessments and fail to simultaneously evaluate the utility of both images and features within a unified framework. To bridge this gap in underwater object detection (one of the typical underwater task), we have developed an underwater image utility dataset and introduced the Lightweight and Unified Utility for Underwater Image Evaluation (LU3IE). LU3IE leverages features derived from multiple detection models to enhance generalization. It then refines these features through dimensionality reduction and commonality extraction, and simplifies complexity through the use of knowledge distillation. Experimental results validate the effectiveness and scalability of LU3IE in accurately assessing the utility of underwater images.

Forward-Looking Sonar (FLS) images play an important role in underwater vision tasks. However, the scarcity of high-quality annotated datasets constrains the performance of data-driven visual models. Due to inherent characteristics such as strong speckle noise, low contrast, and fan-shaped geometric imaging structures, existing natural image generative models are difficult to directly apply to FLS image generation. To address these challenges, we propose SSLDiff, a conditionally controllable sonar image generation framework based on a latent diffusion model. Specifically, a Sonar-Specific Variational Auto-Encoder (SS-VAE) with a tailored loss function is designed to preserve speckle noise statistics and intensity distribution characteristics. Implicit fan-shaped geometric constraints are incorporated during training to model the sonar imaging geometry, and Dual Cross-Attention together with Layout-Masked Attention (LMA) mechanisms enables controllable generation under layout and category conditions. Experimental results demonstrate that the generated images exhibit realistic sonar imaging characteristics and can serve as effective data augmentation, thereby improving downstream object detection performance.

Kinesthetic and tactile information can represent the physical states of objects, encompassing roughness, stiffness, motion, force, and other attributes. The introduction of these can enhance imprecise recognition that relies solely on visual information in cases of light disturbance, occlusion and camouflage. Nevertheless, this task is still challenging due to the heterogeneity among visual, tactile and kinesthetic data. To address this issue, this paper delves into the alignment of heterogeneous data dimensions, the fusion of heterogeneous data features, and the optimization of learning rates for multi-source heterogeneous sensor learning models. Consequently, an effective Visual-Kinesthetic-Tactile Information Fusion (VikitaFusion) network is proposed, which comprises: 1) heterogeneous data extractors that align visual images with tactile and kinesthetic data through image-to-sequence projection; 2) a visual-kinesthetic-tactile Transformer-based domain fusion that mimics human multi-sensory fusion perception through a feature-level fusion block and dynamic fusion blocks; 3) a Periodic Triangulation Learning Rate (PTLR) method aimed at optimizing the learning rate for performance enhancement in multi-source heterogeneous sensor learning models. Extensive experiments demonstrate that VikitaFusion outperforms current state-of-the-art methods with higher recognition accuracy and a lower parameter size.

Sonar images are vital in ocean explorations but face transmission challenges due to limited bandwidth and unstable channels. The Just Noticeable Difference (JND) represents the minimum distortion detectable by human observers. By eliminating perceptual redundancy, JND offers a solution for efficient compression and accurate Image Quality Assessment (IQA) to enable reliable transmission. However, existing JND models prove inadequate for sonar images due to their unique redundancy distributions and the absence of pixel-level annotated data. To bridge these gaps, we propose the first sonar-specific, picture-level JND dataset and a weakly supervised JND model that infers pixel-level JND from picture-level annotations. Our approach starts with pretraining a perceptually lossy/lossless predictor, which collaborates with sonar image properties to drive an unsupervised generator producing Critically Distorted Images (CDIs). These CDIs maximize pixel differences while preserving perceptual fidelity, enabling precise JND map derivation. Furthermore, we systematically investigate JND-guided optimization for sonar image compression and IQA algorithms, demonstrating favorable performance enhancements.

Sonar imaging serves as a crucial sensing modality in underwater environments, yet its outputs are inherently de graded by substantial speckle noise arising from background in terference, biological activity, and device self-noise, posing unique challenges for effective denoising. Existing sonar image denoising techniques take a deterministic approach and produce only one outcome. However, this may neglect the complexities present in real-world situations, thereby limiting the effectiveness of denoising. To address this challenge, we first create a sonar noise dataset with corresponding reference images obtained through well esigned pairwise comparisons. Subsequently, we introduce the Divergence-to-Convergence Sonar Image Denoising (D2C SID) network. This network treats denoising as an uncertainty problem and generates multiple plausible denoised samples. It accomplishes this by leveraging a latent space constructed via a conditional variational autoencode. By this process, our method reduces the reliance on any single estimation. Experimental results on both forward-looking and side-scan sonar images confirm the efficacy of the D2C-SID approach.

Accurate segmentation of cardiac structures in echocardiography is crucial for the early diagnosis of congenital heart disease. Existing methods are limited by the local receptive field of CNNs or the unidirectional scanning of Mamba, resulting in insufficient feature extraction. We propose a multi-directional UMamba network incorporating prototype priors for multi-view segmentation. The network includes a spatio-temporal graph attention module (S-T GAT) to capture dynamic dependencies, four-directional UMamba (FUMamba) for comprehensive structural modeling, and cross-view correlation attention (Cross Attention) to facilitate multi-perspective feature fusion, a prototype manager (Prototype) enhances feature stability under limited annotation. Experiments on multiple datasets demonstrate that our method outperforms other state-of-the-art approaches, achieving a Dice score of 93% on the A4C view and 92% on the PSAX view.

Reliable no-reference image quality assessment (NR-IQA) for sonar imagery remains challenging due to complex acoustic degradations and the limited interpretability of existing models. Most approaches predict a single quality score without explicitly modeling perceptual factors or contextual dependencies. This paper proposes MASI-SIQA, an interpretable NR-IQA framework for sonar images. The method decomposes perceptual quality into four human-aligned attributes, including region discriminability, detail preservation, overall cleanness, and geometric integrity, each modeled by an independent branch to capture heterogeneous degradation characteristics. These attribute representations are integrated with scene-level semantics through a graph-based interaction mechanism for attribute-aware quality prediction. To address the scarcity of annotated sonar data, a transfer learning strategy is adopted to leverage structural representations learned from large-scale natural image datasets and adapt them to sonar-specific characteristics. Experiments are conducted on a self-constructed and publicly available sonar IQA dataset. Quantitative results demonstrate that MASI-SIQA achieves superior performance over state-of-the-art methods in terms of correlation with human perception while providing explicit intermediate representations that enhance interpretability.

Ventricular septal defect (VSD) can be effectively identified by extracting anatomical structures and blood-flow features through multi-scale analysis of multi-modal ultrasound images. However, existing automated identification methods mainly focus on the interaction of multi-scale features, while ignoring the positive impact of hierarchical image differences on identification performance. Moreover, static multi-scale feature extraction strategies within such methods limit the network’s flexibility in modeling features of different frequencies. In this paper, we propose a hierarchical feature-guided dynamic collaborative learning Transformer model for VSD identification. Specifically, we first propose a hierarchical feature extraction unit that sequentially extracts basic- and detailed-layer features from the multi-modal ultrasound images. Subsequently, a hierarchical feature-guided dynamic collaborative learning Transformer block (HDC_Transformer) is constructed. This block performs linear operations while enhancing local anatomical information by using detailed features to guide low-scale features. Simultaneously, a basic-layer feature is used to guide high-scale features for supplementing global information. Moreover, considering that low-scale and high-scale features mainly capture high-frequency and low-frequency information, respectively, the HDC_Transformer block utilizes a dynamic learning mechanism to achieve an effective balance between high- and low-frequency modeling across different feature extraction layers. Extensive experiments on both public and private datasets demonstrate the superior performance of the proposed method across various identification tasks and offer new insights into VSD identification. Our code is available at: https://github.com/Zhaocheng1/HDC_Transformer.

Sonar systems employing underwater acoustic communications serve as critical remote-sensing technologies for deep-ocean monitoring and exploration. However, when transmitted, the sonar images (SIs) are prone to compound degradations at the receiving end, manifested as global blurring, block artifacts, and even information loss. These degradations, caused by compressed data loss and channel-induced deterioration, differ fundamentally from imaging-stage distortions such as seabed reverberation and equipment noise, making existing restoration methods less effective for transmission distortions. To address this issue, we conduct an analysis revealing that transmission distortion predominantly affects the phase component, while the amplitude component contains limited brightness information. Motivated by this observation, we propose a phase-aware contrastive learning framework for sonar image restoration (SIR-PACL) via spatial-frequency fusion. The method employs a frequency-domain branch (FDB) to address phase distortions and integrate phase information into the contrastive learning process to guide more effective representation learning. A spatial-frequency fusion module is further introduced to enhance local detail and edge preservation. Comprehensive evaluations show that SIR-PACL provides improved image restoration quality over state-of-the-art approaches and delivers consistent gains in downstream tasks.

Unmanned underwater vehicles (UUVs) equipped with sensors enable efficient marine resource exploration. However, platform-generated self-noise is inevitably captured by these sensors, degrading the reliability of underwater acoustic communications (UAC). Consequently, separating target signals from noise within the acquired data is critical. Based on data collected from a single-channel hydrophone, this paper proposes two selfnoise suppression methods tailored to classic scenarios: a dataaided approach without a reference source, and a decoupling approach with available reference noise. The effectiveness of the proposed methods is validated through experiments in an anechoic tank. Results demonstrate that the proposed techniques significantly reduce the bit error rate (BER).

Super-resolution technology can improve recognition accuracy of low-resolution face images. However, existing face super-resolution techniques often suffer from identity change and recognizability reduction, limiting their effectiveness in automated face recognition systems. These shortcomings stem from the semantic gap and the lack of high-resolution reference images, leading to suboptimal performance in real-world applications. To reasonably evaluate the performance of face super-resolution, we propose a machine recognition-oriented image quality evaluation method. Firstly, we establish a Machine vision-oriented Face Super-resolution Quality Dataset (M-FSQD) annotated by multiple face recognition algorithms. The dataset is specifically designed to account for machine vision requirements and provides a variety of super-resolution face images for quality assessment. To our knowledge, this is the first face super-resolution image quality dataset tailored for machine vision applications. On this basis, a Siamese network architecture is designed to assess the quality of face super-resolution images quality from a machine vision perspective. Our semi-supervised method effectively extracts the identity-preserving features from low-resolution images while eliminating dependency on high-resolution references. Experimental results demonstrate that our method successfully captures machine-vision-aligned features and outperforms existing image quality assessment methods.

Deep-feature-based perceptual similarity models have demonstrated strong alignment with human visual perception in Image Quality Assessment (IQA). However, most existing approaches operate at a single spatial scale, implicitly assuming that structural similarity at a fixed resolution is sufficient. The role of spatial scale in deep-feature similarity modeling thus remains insufficiently understood. In this letter, we isolate spatial scale as an independent factor using a minimal multiscale extension of DeepSSIM, referred to as Deep Structural Similarity with Multiscale Representation (MSDS). The proposed framework decouples deep feature representation from cross-scale integration by computing DeepSSIM independently across pyramid levels and fusing the resulting scores with a lightweight set of learnable global weights. Experiments on multiple benchmark datasets demonstrate consistent and statistically significant improvements over the single-scale baseline, while introducing negligible additional complexity. The results empirically confirm spatial scale as a non-negligible factor in deep perceptual similarity, isolated here via a minimal testbed.

In recent decades, the emergence of image applications has greatly facilitated the development of vision-based tasks. As a result, image quality assessment (IQA) has become increasingly significant for monitoring, controlling, and improving visual signal quality. While existing IQA methods focus on image fidelity and aesthetics to characterize perceived quality, it is important to evaluate the utility-centered quality of an image for popular tasks, such as object detection. However, research shows that there is a low correlation between utilities and perceptions. To address this issue, this article proposes a utility-centered IQA approach. Specifically, our research focuses on underwater fish detection as a challenging task in an underwater environment. Based on this task, we have developed a utility-centered underwater image quality database (UIQD) and a transfer learning-based advanced underwater quality by utility assessment (AQUA). Inspired by the top-down design approach used in fidelity-oriented IQA methods, we utilize deep models of object detection and transfer their features to the mission of utility-centered quality evaluation. Experimental results validate that the proposed AQUA achieves promising performance not only in fish detection but also in other tasks such as face recognition. We believe that our research provides valuable insights to bridge the gap between IQA research and visual tasks.

Sonar technology has been widely used in underwater surface mapping and remote object detection for its light-independent characteristics. Recently, the booming of artificial intelligence further surges sonar image (SI) processing and understanding techniques. However, the intricate marine environments and diverse nonlinear postprocessing operations may degrade the quality of SIs, impeding accurate interpretation of underwater information. Efficient image quality assessment (IQA) methods are crucial for quality monitoring in sonar imaging and processing. Existing IQA methods overlook the unique characteristics of SIs or focus solely on typical distortions in specific scenarios, which limits their generalization capability. In this article, we propose a unified sonar IQA method, which overcomes the challenges posed by diverse distortions. Though degradation conditions are changeable, ideal SIs consistently require certain properties that must be task-centered and exhibit attribute consistency. We derive a comprehensive set of quality attributes from both the task background and visual content of SIs. These attribute features are represented in just ten dimensions and ultimately mapped to the quality score. To validate the effectiveness of our method, we construct the first comprehensive SI dataset. Experimental results demonstrate the superior performance and robustness of the proposed method.

Sonar imaging system plays a crucial role in ocean exploration since it can overcome the limitations of light conditions. However, the challenge of low resolution remains in sonar images (SIs) due to sonar imaging characteristics and varying compression for low-bandwidth transmission. Most existing image super-resolution (SR) methods treat both the structure and texture in the same way, thus failing to simultaneously capture the rich global-local information. Nevertheless, both structure and texture are essential for the visual quality and applications of SIs. In this study, we propose a structure-texture dual-preserving network (STDPNet) tailored to capture both local texture details and global structure in a parallel manner for SISR. To further explore the internal correlation between structure and texture features, a feature interaction strategy is introduced. Moreover, conventional loss functions for SR often yield smooth results. We propose a hybrid loss function with spectral and local gradient-aware components to preserve frequency content and enhance texture detail. Experimental results validate the superior performance of the proposed STDPNet.

Video compression artifact removal focuses on enhancing the visual quality of compressed videos by mitigating visual distortions. However, existing methods often struggle to effectively capture spatio-temporal features and recover high-frequency details, due to their suboptimal adaptation to the characteristics of compression artifacts. To overcome these limitations, we propose a novel Spatio-Temporal and Frequency Fusion (STFF) framework. STFF incorporates three key components: Feature Extraction and Alignment (FEA), which employs SRU for effective spatiotemporal feature extraction; Bidirectional High-Frequency Enhanced Propagation (BHFEP), which integrates HCAB to restore high-frequency details through bidirectional propagation; and Residual High-Frequency Refinement (RHFR), which further enhances high-frequency information. Extensive experiments demonstrate that STFF achieves superior performance compared to state-of-the-art methods in both objective metrics and subjective visual quality, effectively addressing the challenges posed by video compression artifacts. Trained model available: https://github.com/Stars-WMX/STFF.

Underwater Image Quality Assessment (IQA) provides reliable guidance for the application of underwater image processing. According to the requirements of image processing task, IQAs can be categorized into two classes: perceptionoriented and utility-oriented. Perception-oriented algorithms focus on quality in the human visual system, while utilityoriented algorithms emphasize the application value of images in downstream tasks. Current research reveals that different orientation share universal underlying determinants, enabling the implementation of unified IQA. Thus, we propose a dualoriented IQA framework that simultaneously addresses both perceptual and utility aspects. We explore the universal determinants of IQA under different orientations. For perceptual characteristics, we establish the correlation between underwaterspecific degradations and subjective quality variations to handle diverse underwater environments. For utility characteristics, we employ attention-weighted semantic feature extraction to enhance task relevance of utility representations. Experimental results demonstrate the superior performance and remarkable robustness of our method.