Encoder-decoder Framework Research Articles

Integrating Convolutional Attention and Encoder–Decoder Long Short-Term Memory for Enhanced Soil Moisture Prediction

Soil moisture is recognized as a crucial variable in land–atmosphere interactions. This study introduces the Convolutional Attention Encoder–Decoder Long Short-Term Memory (CAEDLSTM) model to address the uncertainties and limitations inherent in traditional soil moisture prediction methods, especially in capturing complex temporal dynamics across diverse environmental conditions. Unlike existing approaches, this model integrates convolutional layers, an encoder–decoder framework, and multi-head attention mechanisms for the first time in soil moisture prediction. The convolutional layers capture local spatial features, while the encoder–decoder architecture effectively manages temporal dependencies. Additionally, the multi-head attention mechanism enhances the model’s ability to simultaneously focus on multiple key influencing factors, ensuring a comprehensive understanding of complex environmental variables. This synergistic combination significantly improves predictive performance, particularly in challenging climatic conditions. The model was validated using the LandBench1.0 dataset, which includes multiple high-resolution datasets, such as ERA5-land, ERA5 atmospheric variables, and SoilGrids, covering various climatic regions, including high latitudes, temperate zones, and tropical areas. The superior performance of the CAEDLSTM model is evidenced by comparisons with advanced models such as AEDLSTM, CNNLSTM, EDLSTM, and AttLSTM. Relative to the traditional LSTM model, CAEDLSTM achieved an average increase of 5.01% in R2, a 12.89% reduction in RMSE, a 16.67% decrease in bias, and a 4.35% increase in KGE. Moreover, it effectively addresses the limitations of traditional deep learning methods in challenging climates, including tropical Africa, the Tibetan Plateau, and Southeast Asia, resulting in significant enhancements in predictive accuracy within these regions, with R2 values improving by as much as 20%. These results underscore the capabilities of CAEDLSTM in capturing complex soil moisture dynamics, demonstrating its considerable potential for applications in agriculture and water resource monitoring across diverse climates.

Trajectory prediction based on the dynamic characteristics and coupling relationships among vehicles in highway scenarios

In dynamic and highly interactive traffic scenarios, vehicle trajectory prediction has become a critical and complex problem in vehicle assistance systems due to the uncertainty of traffic participant behaviour. This study proposes a novel vehicle trajectory prediction method called the Dynamic Coupling-based Long and Short-Term Memory Network (DSKM-LSTM). The method innovatively combines a vehicle dynamics model with the dynamic coupling between vehicles by integrating the intrinsic motion laws of the vehicles with the correlation of the motion parameters (position, velocity, and acceleration) between the vehicles via Kalman filtering. Subsequently, this information is embedded into a Long Short-Term Memory (LSTM)-based encoder-decoder framework to encode and decode historical trajectories with the fused information in order to predict future driving trajectories. Additionally, this paper introduces a dynamic loss function that optimizes the generalization capability and iteration time of DSKM-LSTM by adjusting the threshold value in real time and dynamically updating the loss value. This approach effectively addresses the problem of insufficient objectivity and accuracy in existing vehicle trajectory prediction models and provides a theoretical foundation for generating objective and interpretable trajectories in the field of trajectory prediction. Validation results on three publicly available datasets show that DSKM-LSTM achieves efficient and accurate long-term (5-s) trajectory prediction in complex traffic scenarios (e.g., motorways).

Correcting translation for non-autoregressive transformer

Non-Autoregressive Transformer has shown great success in recent years. It generally employs the encoder–decoder framework, where the encoder maps the sentence into hidden representation, and the decoder generates the target tokens simultaneously. Since the theory of non-autoregressive transformer is consistent with the architecture of the encoder, we suppose that it is somewhat wasteful for the encoder to only map input sentence into hidden representation. In this study, we proposed a novel non-autoregressive transformer to fully exploit the capabilities of the encoder. Specifically, in our approach, the encoder not only encodes the input sentence into hidden representation, but also generates the target tokens. Consequently, the decoder is relieved of its responsibility to generate the target tokens, instead of focusing on correcting the sentence produced by the encoder. We evaluate the performance of the proposed non-autoregressive transformer on three widely-used translation tasks. The experimental results illustrate the proposed method can significantly improve the performance of the non-autoregressive transformer , which achieved 27.94 BLEU on WMT14 EN → DE task, 33.96 BLEU on WMT16 EN → RO task, and 33.85 BLEU on IWSLT14 DE → EN.

A comprehensive construction of deep neural network‐based encoder–decoder framework for automatic image captioning systems

AbstractThis study introduces a novel encoder–decoder framework based on deep neural networks and provides a thorough investigation into the field of automatic picture captioning systems. The suggested model uses a “long short‐term memory” decoder for word prediction and sentence construction, and a “convolutional neural network” as an encoder that is skilled at object recognition and spatial information retention. The long short‐term memory network functions as a sequence processor, generating a fixed‐length output vector for final predictions, while the VGG‐19 model is utilized as an image feature extractor. For both training and testing, the study uses a variety of photos from open‐access datasets, such as Flickr8k, Flickr30k, and MS COCO. The Python platform is used for implementation, with Keras and TensorFlow as backends. The experimental findings, which were assessed using the “bilingual evaluation understudy” metric, demonstrate the effectiveness of the suggested methodology in automatically captioning images. By addressing spatial relationships in images and producing logical, contextually relevant captions, the paper advances image captioning technology. Insightful ideas for future study directions are generated by the discussion of the difficulties faced during the experimentation phase. By establishing a strong neural network architecture for automatic picture captioning, this study creates opportunities for future advancement and improvement in the area.

Can question-texts improve the recognition of handwritten mathematical expressions in respondents’ solutions?

The accurate recognition of respondents’ handwritten solutions is important for implementing intelligent diagnosis and tutoring. This task is significantly challenging because of scribbled and irregular writing, especially when handling primary or secondary students whose handwriting has not yet been fully developed. Recognition becomes difficult in such cases even for humans relying only on the visual signals of handwritten content without any context. However, despite decades of work on handwriting recognition, few studies have explored the idea of utilizing external information (question priors) to improve the accuracy. Based on the correlation between questions and solutions, this study aims to explore whether question-texts can improve the recognition of handwritten mathematical expressions (HMEs) in respondents’ solutions. Based on the encoder–decoder framework, which is the mainstream method for HME recognition, we propose two models for fusing question-text signals and handwriting-vision signals at the encoder and decoder stages, respectively. The first, called encoder-fusion, adopts a static query to implement the interaction between two modalities at the encoder phase, and to better catch and interpret the interaction, a fusing method based on a dynamic query at the decoder stage, called decoder-attend is proposed. These two models were evaluated on a self-collected dataset comprising approximately 7k samples and achieved accuracies of 62.61% and 64.20%, respectively, at the expression level. The experimental results demonstrated that both models outperformed the baseline model, which utilized only visual information. The encoder fusion achieved results similar to those of other state-of-the-art methods.

Radar Echo Extrapolation Based on Translator Coding and Decoding Conditional Generation Adversarial Network

In response to the shortcomings of current spatiotemporal prediction models, which frequently encounter difficulties in temporal feature extraction and the forecasting of medium to high echo intensity regions over extended sequences, this study presents a novel model for radar echo extrapolation that combines a translator encoder-decoder architecture with a spatiotemporal dual-discriminator conditional generative adversarial network (STD-TranslatorNet). Initially, an image reconstruction network is established as the generator, employing a combination of a temporal attention unit (TAU) and an encoder–decoder framework. Within this architecture, both intra-frame static attention and inter-frame dynamic attention mechanisms are utilized to derive attention weights across image channels, thereby effectively capturing the temporal evolution of time series images. This approach enhances the network’s capacity to comprehend local spatial features alongside global temporal dynamics. The encoder–decoder configuration further bolsters the network’s proficiency in feature extraction through image reconstruction. Subsequently, the spatiotemporal dual discriminator is crafted to encapsulate both temporal correlations and spatial attributes within the generated image sequences. This design serves to effectively steer the generator’s output, thereby augmenting the realism of the produced images. Lastly, a composite multi-loss function is proposed to enhance the network’s capability to model intricate spatiotemporal evolving radar echo data, facilitating a more comprehensive assessment of the quality of the generated images, which in turn fortifies the network’s robustness. Experimental findings derived from the standard radar echo dataset (SRAD) reveal that the proposed radar echo extrapolation technique exhibits superior performance, with average critical success index (CSI) and probability of detection (POD) metrics per frame increasing by 6.9% and 7.6%, respectively, in comparison to prior methodologies.

Learning Omni-Dimensional Spatio-Temporal Dependencies for Millimeter-Wave Radar Perception

Reliable environmental perception capabilities are a prerequisite for achieving autonomous driving. Cameras and LiDAR are sensitive to illumination and weather conditions, while millimeter-wave radar avoids these issues. Existing models rely heavily on image-based approaches, which may not be able to fully characterize radar sensor data or efficiently further utilize them for perception tasks. This paper rethinks the approach to modeling radar signals and proposes a novel U-shaped multilayer perceptron network (U-MLPNet) that aims to enhance the learning of omni-dimensional spatio-temporal dependencies. Our method involves innovative signal processing techniques, including a 3D CNN for spatio-temporal feature extraction and an encoder–decoder framework with cross-shaped receptive fields specifically designed to capture the sparse and non-uniform characteristics of radar signals. We conducted extensive experiments using a diverse dataset of urban driving scenarios to characterize the sensor’s performance in multi-view semantic segmentation and object detection tasks. Experiments showed that U-MLPNet achieves competitive performance against state-of-the-art (SOTA) methods, improving the mAP by 3.0% and mDice by 2.7% in RD segmentation and AR and AP by 1.77% and 2.03%, respectively, in object detection. These improvements signify an advancement in radar-based perception for autonomous vehicles, potentially enhancing their reliability and safety across diverse driving conditions.

EDH-STNet: An Evaporation Duct Height Spatiotemporal Prediction Model Based on Swin-Unet Integrating Multiple Environmental Information Sources

Given the significant spatial non-uniformity of marine evaporation ducts, accurately predicting the regional distribution of evaporation duct height (EDH) is crucial for ensuring the stable operation of radio systems. While machine-learning-based EDH prediction models have been extensively developed, they fail to provide the EDH distribution over large-scale regions in practical applications. To address this limitation, we have developed a novel spatiotemporal prediction model for EDH that integrates multiple environmental information sources, termed the EDH Spatiotemporal Network (EDH-STNet). This model is based on the Swin-Unet architecture, employing an Encoder–Decoder framework that utilizes consecutive Swin-Transformers. This design effectively captures complex spatial correlations and temporal characteristics. The EDH-STNet model also incorporates nonlinear relationships between various hydrometeorological parameters (HMPs) and EDH. In contrast to existing models, it introduces multiple HMPs to enhance these relationships. By adopting a data-driven approach that integrates these HMPs as prior information, the accuracy and reliability of spatiotemporal predictions are significantly improved. Comprehensive testing and evaluation demonstrate that the EDH-STNet model, which merges an advanced deep learning algorithm with multiple HMPs, yields accurate predictions of EDH for both immediate and future timeframes. This development offers a novel solution to ensure the stable operation of radio systems.

Classification-Based Parameter Optimization Approach of the Turning Process

The turning process is a widely used machining process, and its productivity has a significant impact on the cost and profit in industrial enterprises. Currently, it is difficult to effectively determine the optimum process parameters under complex conditions. To address this issue, a classification-based parameter optimization approach of the turning process is proposed in this paper, which aims to provide feasible optimization suggestions of process parameters and consists of a classification model and several optimization strategies. Specifically, the classification model is used to separate the whole complex process into different substages to reduce difficulties of the further optimization, and it achieves high accuracy and strong anti-interference in the identification of substages by integrating the advantages of an encoder-decoder framework, attention mechanism, and major voting. Additionally, during the optimization process of each substage, Dynamic Time Warping (DTW) and K-Nearest Neighbor (KNN) are utilized to eliminate the negative impact of cutting tool wear status on optimization results at first. Then, the envelope curve strategy and boxplot method succeed in the adaptive calculation of a parameter threshold and the detection of optimizable items. According to these optimization strategies, the proposed approach performs well in the provision of effective optimization suggestions. Ultimately, the proposed approach is verified by a bearing production line. Experimental results demonstrate that the proposed approach achieves a significant productivity improvement of 23.43% in the studied production line.

Research on stock prediction based on CED-PSO-StockNet time series model

To tackle the challenge of low accuracy in stock prediction within high-noise environments, this paper innovatively introduces the CED-PSO-StockNet time series model. Initially, the model decomposes raw stock data using the Complete Ensemble Empirical Mode Decomposition with Adaptive Noise (CEEMDAN) technique and reconstructs the components by estimating their frequencies via the extreme point method. This process enhances component stability and mitigates noise interference. Subsequently, an Encoder-Decoder framework equipped with an attention mechanism is employed for precise prediction of the reconstructed components, facilitating more effective extraction and utilization of data features. Furthermore, this paper utilizes an Improved Particle Swarm Optimization (IPSO) algorithm to optimize the model parameters. On the Pudong Bank dataset, through ablation experiments and comparisons with baseline models, various optimization strategies incorporated into the proposed CED-PSO-StockNet model were effectively validated. Compared to the standalone LSTM model, CED-PSO-StockNet achieved a remarkable 45.59% improvement in the R2 metric. To further assess the model’s generalization capability, this paper also conducted comparative experiments on the Ping An Bank dataset, and the results underscored the significant advantages of CED-PSO-StockNet in the domain of stock prediction.

MFHARFNet: multi-branch feature hybrid and adaptive receptive field network for image segmentation

Abstract Accurate segmentation of medical images is crucial for disease diagnosis and understanding disease changes. Deep learning methods, utilizing encoder-decoder structures, have demonstrated cutting-edge performance in various medical image segmentation tasks. However, the pooling operation in the encoding stage results in feature loss, which makes the network lack the ability to fuse multi-scale information at different levels, hinders its effective perception of multi-scale information, and leads to poor segmentation performance. Drawing inspiration from the U-shaped network, this study introduces a multi-branch feature hybrid attention and adaptive receptive field network (MFHARFNet) for medical image segmentation. Building upon the encoder-decoder framework, we initially devise a multi-branch feature hybrid attention module (MFHAM) to seamlessly integrate feature maps of varying scales, capturing both fine-grained features and coarse-grained semantics across the entire scale. Furthermore, we redesign the skip connection to amalgamate feature information from different branches in the encoder stage and efficiently transmit it to the decoder, providing the decoder with global context feature maps at different levels. Finally, the adaptive receptive field (ARF) module is introduced in the decoder feature reconstruction stage to adapt and focus on related fields, ensuring the model’s adaptation to different segmentation target features, and achieving different weights for the output of different convolution kernels to improve segmentation performance. We comprehensively evaluate our method on medical image segmentation tasks, by using four public datasets across CT and MRI. Remarkably, MFHARFNet method consistently outperforms other state-of-the-art methods, exceeding UNet by 2.1%, 0.9%, 6.6% and 1.0% on Dice on ATLAS, LiTs, BraTs2019 and Spine and intervertebral disc datasets, respectively. In addition, MFHARFNet minimizes network parameters and computational complexity as much as possible. The source codes are in https://github.com/OneHundred99/MFHARFNet.

A Review of Deep Learning-Based Remote Sensing Image Caption: Methods, Models, Comparisons and Future Directions

Remote sensing images contain a wealth of Earth-observation information. Efficient extraction and application of hidden knowledge from these images will greatly promote the development of resource and environment monitoring, urban planning and other related fields. Remote sensing image caption (RSIC) involves obtaining textual descriptions from remote sensing images through accurately capturing and describing the semantic-level relationships between objects and attributes in the images. However, there is currently no comprehensive review summarizing the progress in RSIC based on deep learning. After defining the scope of the papers to be discussed and summarizing them all, the paper begins by providing a comprehensive review of the recent advancements in RSIC, covering six key aspects: encoder–decoder framework, attention mechanism, reinforcement learning, learning with auxiliary task, large visual language models and few-shot learning. Subsequently a brief explanation on the datasets and evaluation metrics for RSIC is given. Furthermore, we compare and analyze the results of the latest models and the pros and cons of different deep learning methods. Lastly, future directions of RSIC are suggested. The primary objective of this review is to offer researchers a more profound understanding of RSIC.

Transformer with token attention and attribute prediction for image captioning

Recently, Vision Transformers (ViTs) have become the mainstream models in image captioning tasks. ViTs take all image tokens as inputs to extract visual features, which may cause concerns about worthless tokens, and meanwhile lead to a huge amount of computation. This paper proposes a novel token reduction module to remedy this drawback. Specifically, the module employs ViTs to embed the input tokens, and adaptively learns informative visual tokens in way of token attention on the channel-spatial granularity. Furthermore, an attribute prediction module is designed to strengthen the relationship between vision and language. Technically, the attribute prediction is achieved via a classifier in form of Multi-Layer Perceptron (MLP). Both the visual representations and attribute representations are obtained by Transformers, which are then combined as the input of the Transformer decoder for caption generation. All of the modules are constructed in an encoder–decoder framework and support the end-to-end learning. Experiment results have shown that our approach can effectively reduce the computational cost of ViTs while maintaining comparable performance on the MS COCO and NoCaps datasets. For example, by pruning more than 70% of the input tokens, our approach greatly reduces GFLOPs by 41% ∼ 47%, while preserving its accuracy of a 142.1 CIDEr score on the MS COCO dataset.

Deep learning aided underwater acoustic OFDM receivers: Model-driven or data-driven?

The Underwater Acoustic (UWA) channel is bandwidth-constrained and experiences doubly selective fading. It is challenging to acquire perfect channel knowledge for Orthogonal Frequency Division Multiplexing (OFDM) communications using a finite number of pilots. On the other hand, Deep Learning (DL) approaches have been very successful in wireless OFDM communications. However, whether they will work underwater is still a mystery. For the first time, this paper compares two categories of DL-based UWA OFDM receivers: the Data-Driven (DD) method, which performs as an end-to-end black box, and the Model-Driven (MD) method, also known as the model-based data-driven method, which combines DL and expert OFDM receiver knowledge. The encoder-decoder framework and Convolutional Neural Network (CNN) structure are employed to establish the DD receiver. On the other hand, an unfolding-based Minimum Mean Square Error (MMSE) structure is adopted for the MD receiver. We analyze the characteristics of different receivers by Monte Carlo simulations under diverse communications conditions and propose a strategy for selecting a proper receiver under different communication scenarios. Field trials in the pool and sea are also conducted to verify the feasibility and advantages of the DL receivers. It is observed that DL receivers perform better than conventional receivers in terms of bit error rate.

AL-SAR: Active Learning for Skeleton-Based Action Recognition.

Action recognition from temporal multivariate sequences of features, such as identifying human actions, is typically approached by supervised training as it requires many ground truth annotations to reach high recognition accuracy. Unsupervised methods for the organization of sequences into clusters have been introduced, however, such methods continue to require annotations to associate clusters with actions. The challenges in annotation necessitate an effective classification methodology that minimizes the required number of labels. Active learning (AL) approaches have been proposed to address these challenges and were able to establish robust results on image classification. Such approaches are not directly applicable to sequences, since for sequences, the variations are in both spatial and temporal domains. In this brief, we introduce a novel method for AL for sequences, called "AL-SAR," which combines unsupervised training with sparsely supervised annotation. In particular, AL-SAR employs a multi-head mechanism for robust uncertainty evaluation of the latent space learned by an encoder-decoder framework. It aims to iteratively select a sparse set of samples, which annotation contributes the most to the disentanglement of the latent space. We evaluate our system on common benchmark datasets with multiple sequences and actions, such as NW-UCLA, NTU RGB+D 60, and UWA3D. Our results indicate that AL-SAR coupled with encoder-decoder network outperforms other AL methods coupled with the same network structure.

Probabilistic Keyphrase Generation From Copy and Generating Spaces.

Keyphrase generation is one of the most fundamental tasks in natural language processing (NLP). Most existing works on keyphrase generation mainly focus on using holistic distribution to optimize the negative log-likelihood loss, but they do not directly manipulate the copy and generating spaces, which may reduce the generability of the decoder. Additionally, existing keyphrase models are either unable to determine the dynamic numbers of keyphrases or produce the number of keyphrases implicitly. In this article, we propose a probabilistic keyphrase generation model from copy and generating spaces. The proposed model is built upon the vanilla variational encoder-decoder (VED) framework. On top of VED, two separate latent variables are adopted to model the distribution of data within the latent copy and generating spaces, respectively. Specifically, we adopt a von Mises-Fisher (vMF) distribution to obtain a condensed variable for modifying the generating probability distribution over the predefined vocabulary. Meanwhile, we utilize a clustering module, which is designed to promote Gaussian Mixture learning and subsequently extract a latent variable for the copy probability distribution. Moreover, we utilize a natural property of the Gaussian mixture network and use the number of filtered components to determine the number of keyphrases. The approach is trained based on latent variable probabilistic modeling, neural variational inference, and self-supervised learning. Experiments on social media and scientific article datasets outperform the state-of-the-art baselines in generating accurate predictions and controllable keyphrase numbers.

IDM-Net: A Multi-Task Supported Encoder-Decoder Framework for Magnetic Field Inverse Design

IDM-Net: A Multi-Task Supported Encoder-Decoder Framework for Magnetic Field Inverse Design

FCSwinU: Fourier Convolutions and Swin Transformer UNet for Hyperspectral and Multispectral Image Fusion.

The fusion of low-resolution hyperspectral images (LR-HSI) with high-resolution multispectral images (HR-MSI) provides a cost-effective approach to obtaining high-resolution hyperspectral images (HR-HSI). Existing methods primarily based on convolutional neural networks (CNNs) struggle to capture global features and do not adequately address the significant scale and spectral resolution differences between LR-HSI and HR-MSI. To tackle these challenges, our novel FCSwinU network leverages the spectral fast Fourier convolution (SFFC) module for spectral feature extraction and utilizes the Swin Transformer's self-attention mechanism for multi-scale global feature fusion. FCSwinU employs a UNet-like encoder-decoder framework to effectively merge spatiospectral features. The encoder integrates the Swin Transformer feature abstraction module (SwinTFAM) to encode pixel correlations and perform multi-scale transformations, facilitating the adaptive fusion of hyperspectral and multispectral data. The decoder then employs the Swin Transformer feature reconstruction module (SwinTFRM) to reconstruct the fused features, restoring the original image dimensions and ensuring the precise recovery of spatial and spectral details. Experimental results from three benchmark datasets and a real-world dataset robustly validate the superior performance of our method in both visual representation and quantitative assessment compared to existing fusion methods.

A Novel RUL-Centric Data Augmentation Method for Predicting the Remaining Useful Life of Bearings

Maintaining the reliability of rotating machinery in industrial environments entails significant challenges. The objective of this paper is to develop a methodology that can accurately predict the condition of rotating machinery in order to facilitate the implementation of effective preventive maintenance strategies. This article proposed a novel RUL-centric data augmentation method, designated as DF-MDAGRU, for the purpose of predicting the remaining useful life (RUL) of bearings. This model is based on an encoder–decoder framework that integrates time–frequency domain feature enhancement with multidimensional dynamic attention gated recurrent units for feature extraction. This method enhances time–frequency domain features through the Discrete Wavelet Downsampling module (DWD) and Convolutional Fourier Residual Block (CFRB). This method employs a Multiscale Channel Attention Module (MS-CAM) and a Multiscale Convolutional Spatial Attention Mechanism (MSSAM) to extract channel and spatial feature information. Finally, the output predictions are processed through linear regression to achieve the final RUL estimation. Experimental results demonstrate that the proposed method outperforms other state-of-the-art approaches on the FEMETO-ST and XJTU datasets.

Dual Stream Encoder–Decoder Architecture with Feature Fusion Model for Underwater Object Detection

Underwater surveillance is an imminent and fascinating exploratory domain, particularly in monitoring aquatic ecosystems. This field offers valuable insights into underwater behavior and activities, which have broad applications across various domains. Specifically, underwater surveillance involves detecting and tracking moving objects within aquatic environments. However, the complex properties of water make object detection a challenging task. Background subtraction is a commonly employed technique for detecting local changes in video scenes by segmenting images into the background and foreground to isolate the object of interest. Within this context, we propose an innovative dual-stream encoder–decoder framework based on the VGG-16 and ResNet-50 models for detecting moving objects in underwater frames. The network includes a feature fusion module that effectively extracts multiple-level features. Using a limited set of images and performing training in an end-to-end manner, the proposed framework yields accurate results without post-processing. The efficacy of the proposed technique is confirmed through visual and quantitative comparisons with eight cutting-edge methods using two standard databases. The first one employed in our experiments is the Underwater Change Detection Dataset, which includes five challenges, each challenge comprising approximately 1000 frames. The categories in this dataset were recorded under various underwater conditions. The second dataset used for practical analysis is the Fish4Knowledge dataset, where we considered five challenges. Each category, recorded in different aquatic settings, contains a varying number of frames, typically exceeding 1000 per category. Our proposed method surpasses all methods used for comparison by attaining an average F-measure of 0.98 on the Underwater Change Detection Dataset and 0.89 on the Fish4Knowledge dataset.