We have developed and evaluated four distinct methodologies for CLFP recognition. The overall architecture of our proposed framework is depicted in Figure 1, which outlines the integrated pipeline for feature extraction, optimization, clustering, and classification. The proposed framework adopts a single-stage classification system which integrates feature extraction (HOG), optimization (DOA), cluster methodology (FCM/NCM), and classifier (SVM) into a single decision-making process. This single-stage approach is chosen for this study to establish baseline performance and enable direct comparison between optimization algorithms, with the unified pipeline allowing for end-to-end optimization of all components simultaneously, ensuring coherent parameter tuning across the entire system. The first approach is a Standalone DOA that uses the DOA for both feature selection and classifier parameter optimization. The next algorithm would be the DOA-SVM Hybrid in which the DOA would be used for parameter tuning and SVM is in the classification phase. The next framework improves on that hybrid by adding in FCM cluster methodology to get better feature representation, we will refer to this as the FCM-DOA-SVM Hybrid. Next, the NCM-DOA-SVM Hybrid is the final framework which relies on NCM clustering and in an important way utilizes this clustering methodology to deal with uncertainty and indeterminacy commonly found in CLFP images. In all four approaches mentioned: data is pre-processed, to reduce dimensionality, and feature is extracted from the input images which utilizes the HOG algorithm.

Preprocessing is a critical first step that directly impacts recognition accuracy by standardizing image quality and preparing the data for subsequent analysis. Our preprocessing pipeline consists of two key steps. Step 1 involves resizing the image to a fixed dimension of 128 × 128 pixels to ensure uniformity across all input samples. Step 2 entails converting the image from RGB (color) format to grayscale. Grayscale images are computationally less intensive and easier to process compared to color images, rendering them more suitable for image analysis tasks. These preprocessing steps are implemented using OpenCV in Python, specifically employing the cv2.resize() function for resizing and cv2.cvtColor() with the cv2.COLOR_RGB2GRAY flag for color-to-grayscale conversion.

A significant issue encountered by excellence algorithms is the presence of redundant and superfluous attributes, which lead to an increase in data volume and consequently expand the research space needed to address the problem. This expansion results in prolonged processing time for data necessary for detection and classification (Soranamageswari and Meena, 2010; Vithlani and Kumbharana, 2015). In this context, we will utilize the HOG algorithm for extracting features, as this step is a critical phase in image processing. The HOG algorithm will be employed to extract significant features from images, thereby facilitating the classification process and reducing data size, which in turn decreases the time required for classification (Moen, 2018; Cetina et al., 2014).

The fundamental stages of the HOG algorithm (Elhariri et al., 2015) are outlined as follows:

Figure 3 presents a schematic representation of the utilization of the HOG algorithm for extracting features from CLFP images.

In this research, traits were extracted by identifying the most critical features that encapsulate the maximum amount of information from the dataset. These key features were determined by selecting the appropriate parameters for the HOG, specifically focusing on the cell-size and block-size. The implementation of the HOG technique was executed in Python utilizing the skimage.feature library, with the hog function.

To ensure robust model performance and prevent overfitting, we implemented a comprehensive validation framework within our existing train-test split. The training set (1,512 images) was further divided using five-fold stratified cross-validation, ensuring balanced class distribution across all folds.

DOA is a new method that gets its ideas from how dolphins communicate and hunt cooperatively (Wu et al., 2016; Al-Taie and Khaleel, 2024). Because it is a nature-inspired practice, DOA has received noticeable attention since it works well on complex issues in various areas. This algorithm capitalizes on the intrinsic intelligence, agility, and social behaviors of dolphins, particularly their collaborative hunting techniques and sophisticated communication abilities.

Some of the traits show that dolphins are recognized as one of the most intelligent marine species, such as echo-location, team cooperation, task allocation, and acoustic communication. These distinctive traits have been effectively integrated into the DOA allowing to look for new solutions and still improve existing ones while managing to use resources wisely (Gholizadeh and Poorhoseini, 2016).

DOA copies how dolphins in nature move together which allows for better navigation through complicated and wide-ranging problem spaces (Soto et al., 2016; Yong et al., 2016). The main elements of the algorithm are a location update method, an equation for speed changes and using echolocation to find the best solutions (Wu et al., 2017; Kaveh et al., 2017). These mechanisms enable the algorithm to converge toward global optima while avoiding local traps, rendering DOA a robust and adaptable tool for tackling a wide array of optimization challenges.

ML specifically supervised learning algorithms, is employed in both regression and classification problems (Faye et al., 2018). Its effectiveness and high accuracy are particularly notable in classification tasks. In the context of the SVM classifier, the selection of optimal parameters is crucial for achieving high performance in biometric identification systems (Qi et al., 2013; Shao et al., 2011). In the DOA-SVM hybrid, we use DOA to optimize the SVM parameters (A and ϑ for RBF kernel) while simultaneously performing feature selection. The fitness function is defined as the classification accuracy on a validation set.

Using the DOA and SVM methods simultaneously provides a modern approach to improve solving the classification problems with the advantages of both approaches. Such algorithms can interact and use their smartness together to enhance prediction accuracy and ensure the best possible outcomes which improves the model's overall performance. Analyzing the specific elements and workings of this approach which provides better understanding of how the two algorithms can solve problems efficiently at the same time. Also, by looking into the problems and possible improvements of these integration techniques, the ways to improve optimization and classification methods in the future can be determined, as demonstrated in Algorithm 2.

Clustering algorithms such as FCM are extensively employed in preprocessing tasks to identify significant patterns and structures within data. FCM is especially useful in soft clustering contexts, where each data point can be associated with multiple clusters to varying extents (Mendel and Bonissone, 2021; Chimatapu et al., 2018). This capability is advantageous in biometric applications, where data frequently displays overlapping features. In the FCM-DOA-SVM hybrid model, the FCM algorithm improves the input features by adding fuzzy membership values to each data sample, thus capturing essential structural information prior to classification (Shukla et al., 2020; Ferreyra et al., 2019).

Subsequently, the DOA is utilized to fine-tune the parameters of the SVM classifier, specifically (A and ϑ for the RBF kernel). This integrated method effectively merges the unsupervised clustering capability of FCM with the supervised learning of SVM and the global optimization potential of DOA. In this scenario, the fitness function is characterized by the classification accuracy achieved on a validation set using the SVM trained with features enhanced by FCM.

The FCM-DOA-SVM architecture exhibits a significant synergistic effect, wherein fuzzy memberships enhance feature representation, and DOA facilitates the exploration of optimal classifier configurations. Consequently, the model demonstrates enhanced generalization and robustness in biometric classification tasks, as detailed in Algorithm 3.

NCM-DOA-SVM hybrid includes the NCM method of clustering which is helpful for CLFP recognition to handle indeterminacy in images. Neutrosophic Theory represents the information that is uncertain, imprecise and inconsistent by using three degrees, truth (T), indeterminacy (I) and falsity (F). The NCM clustering algorithm extends the classical FCM by incorporating this neutrosophic concept, enabling more accurate modeling of real-world biometric data. In the proposed NCM-DOA-SVM hybrid, T, I and F are extracted for every sample during the clustering procedure which enhance feature representations.

Subsequently, the DOA algorithm is used to improve the hyperparameters of the SVM classifier, specifically the penalty value (A) and the RBF kernel value (ϑ). Using this hybrid approach boosts the classifier's results by finding the optimal parameters and additionally making use of the extra features that capture the neutrosophic nature of the input data.

The integration of NCM for handling unclear data, DOA for finding global solution and SVM for accurate classification leads to the development of a strong model for biometric recognition. By applying the combined model, it deals with uncertainty and the classifier becomes more effective as shown in Algorithm 4.

At this stage, a dataset consisting of 2016 CLFP images was used to develop the DOA and the suggested hybrid approaches, namely DOA-SVM, FCM-DOA-SVM, and NCM-DOA-SVM. The HOG technique was used for initial image processing and trait extraction, which made it easier to extract a feature matrix from these images. The DOA and the suggested hybrid approaches (DOA-SVM, FCM-DOA-SVM, and FCM-DOA-SVM) were then trained using the derived traits matrix, improving their performance.

Upon the completion of the training phase and achieving algorithm stability, a test was conducted involving 504 CLFP images. These images were subjected to preprocessing, followed by feature extraction using the HOG algorithm. To determine the optimal HOG feature extraction parameters, we conducted a comprehensive sensitivity analysis across different cell and block size configurations.

As shown in Table 3, three configurations were evaluated: Fine (16 × 16 cells, 2 × 2 blocks), Medium (32 × 32 cells, 2 × 2 blocks), and Coarse (64 × 64 cells, 2 × 2 blocks). The results demonstrate a clear trade-off between feature dimensionality, computational efficiency, and classification accuracy. The Medium configuration achieved the highest accuracy of 0.856198, despite having a moderate feature vector size of 324 dimensions. Interestingly, the Fine configuration, with its substantially larger feature vector (1,764 dimensions), yielded lower accuracy (0.844628), suggesting that excessive granularity may introduce noise or lead to overfitting. The negative correlation between feature size and accuracy (r = −0.2498) supports this observation. Statistical analysis using ANOVA revealed no significant difference between configurations (p≥0.05), indicating that the Medium configuration provides an optimal balance between computational efficiency (0.003086) and classification performance. The choice of 32 × 32 cell sizes, resulting in four cells for a 128 × 128 image, represents a strategic compromise that captures sufficient local gradient information while maintaining computational tractability and avoiding the curse of dimensionality associated with finer granularities.

The assignment of specific coefficients to each algorithm is determined by the nature of the task, as the specific operations may vary based on the research problem being addressed. Through a series of practical experiments and their application, the appropriate coefficients for the algorithms were established. The values of these coefficients are detailed in Table 4.

The dataset which included 1,512 images, was processed using the DOA method to identify CLFP. The following results were obtained from the evaluation of the suggested algorithm's training using predetermined evaluation criteria:

The dataset include 504 images, which was assessed using the DOA system to determine CLFP. Table 5 provides details on the training results of the proposed algorithm.

The results of the test phase of DOA are detailed in Table 5, which looks at different configurations of the coefficients: cell-size HOG is always set at 32, block-size HOG is fixed at 2, dolphin population level remains at 30, and iteration threshold is investigated at 50.

The DOA-SVM system was employed to train a dataset comprising 1,512 CLFP images for the purpose of CLFP identification. The performance of the proposed hybrid algorithm was evaluated using established assessment criteria, yielding the following results:

The DOA-SVM system was used to analyze the dataset, which included 504 images, in order to find CLFP. In Table 6, the results are presented in detail.

The results of the test phase of the hybrid method DOA-SVM, which examines different coefficient configurations, are detailed in Table 6. HOG with a block-size fixed at 2, a dolphin population level held constant at 30, an iteration threshold examined at 50, and a cell-size continuously maintained at 32.

The FCM-DOA-SVM system was trained to identify CLFP using a dataset comprising 1,512 images. The effectiveness of the proposed algorithm was assessed using evaluation metrics and these are its outcomes:

The data was assessed utilizing the FCM-DOA-SVM system to identify CLFP, comprising 504 images. The output of the algorithm training is shown in Table 7 (i.e., the test phase results of the proposed hybrid method are shown in this table). The FCM-DOA-SVM setup kept HOG cell-size at 32, HOG block-size at 2 and Dolphin population at 30 and it varied the Iteration threshold from 1 to 50.

The NCM-DOA-SVM model was trained on a dataset of 1,512 CLFP images, and its performance was assessed using standard evaluation metrics. The results are as given below:

In order to analyze and identify the 504 images in the dataset as CLFP, the NCM-DOA-SVM system has been used. Table 8 lists the training phase results and provides an explanation of the testing phase results using the suggested hybrid approach. The following parameter settings were used when implementing the NCM-DOA-SVM system: The dolphin population level was kept at 30, the iteration threshold was investigated at 50, and the HOG cell-size was continuously set at 32 and the HOG block-size was fixed at 2.

The rating ratio rate is shown in Table 9 for the DOA method, as well as for the proposed hybrid approaches: DOA-SVM, FCM-DOA-SVM, and NCM-DOA-SVM.

The experimental results confirmed the robust performance of all proposed algorithms during the training phase. The best classification accuracy, 91.0%, was found for images with a size of 128 × 128 and a cell-size of 32 × 32, as shown in Table 5. Table 6 showed an improved classification accuracy of 94.07% for images of the same dimensions and cell-size when the number of dolphins matched the number of repetitions. Additionally, Tables 7, 8 reported a further increase in classification accuracy to 96.03 and 98.0% under the same conditions. These findings are corroborated by the data in Table 9. The DOA showed an average accuracy rate of 91.0%, whereas the DOA-SVM hybrid algorithm achieved an average accuracy of 94.07%, the FCM-DOA-SVM algorithm attained an average accuracy of 96.03% and NCM-DOA-SVM algorithm reached an average accuracy of 98.0%.

Our experimental findings indicate that the hybrid algorithm DOA-SVM demonstrates a higher classification ratio than the DOA. This improvement is attributed to the DOA optimization capabilities, which emulate the cooperative behavior of dolphins to enhance parameter selection. The algorithm improves convergence speed and reduces computational cost by learning from previous configurations. A well-tuned DOA balances exploration and exploitation, making parameter selection crucial for optimal performance.

According to experimental results, the suggested hybrid algorithm, which combines DOA-SVM with FCM, performs better in terms of classification accuracy than both the standalone DOA and the hybrid DOA-SVM algorithm. The efficient combination of FCM and DOA, where each technique improves the performance of the other, is the source of this improvement. When these algorithms are used together, they perform better on classification tasks and other tasks than when they are used separately. The FCM algorithm helps by providing soft aggregation capabilities, which enable more precise data point classification, particularly when features overlap. The DOA also supports more efficient searching, so that key data points can be identified and the required computations can be reduced.

Within the spectrum of assessed approaches, the NCM-DOA-SVM method consistently maintained superior classification accuracy. This enhanced performance is resulted due to the effective combination of NCM clustering, the DOA, and the SVM. This approach is convenient for handling unclear, overlapping and uncertain segments within the data and this quality is key for classifying complicated data sets. On the other hand, the DOA helps with global optimization, so features can be effectively chosen and parameters can be successfully tuned. Finally, the SVM provides good performance in the final classification stage. The integration of all three methods together creates a better and more efficient results than if only two are used, such as DOA or DOA-SVM.

In this neutrosophic logic visualization (see Figure 4), the T/I/F membership framework provides significant advantages for our database. Specifically, Truth membership (T) highlights robust and well-defined clusters at the core of the dataset, while Indeterminacy membership (I) captures transitional and boundary regions–effectively representing ambiguous cases that are often misclassified by traditional methods. Falsity membership (F) excels at identifying outliers, enhancing data integrity by revealing non-conforming patterns that may compromise analysis. The integrated T/I/F analysis enables a comprehensive and nuanced characterization of uncertainty, offering deeper insights and improved decision-making accuracy over conventional single-value approaches, thus establishing neutrosophic logic as a superior tool for extracting actionable information from uncertain databases.

In this work, we propose a CLFP recognition system based on the DOA with 3,902.533 processing seconds and 91.0% classification accuracy. Additionally, we combine the DOA and the SVM algorithms to propose a new hybrid algorithm. By suggesting a new fitness function based on the SVM algorithm that depends on dolphin values rather than the original parameters (G, ϵ, A), the combination aims to improve classification accuracy. This method took 5,039.768 s to process and had a classification accuracy of 94.07%. In order to further increase classification accuracy, we also suggest another hybrid algorithm that combines the DOA-SVM and the FCM algorithm. In comparison to both the DOA and the DOA-SVM algorithm, this algorithm efficiently handles ambiguous and imprecise data, resulting in a more accurate identification system and attaining classification accuracy of 96.03% with a processing time of 1,684.920 s. We suggest combining NCM with DOA-SVM in place of FCM. This results in a higher classification accuracy of 98.0% at a processing time of 1,352.024 s. The NCM-DOA-SVM outperforms other methods for CLFP images.

The comparative evaluation of bio-inspired optimization algorithms (see Table 10) reveals that DOA-SVM achieves the highest classification accuracy of 89.69% with the lowest standard deviation (±0.0132), demonstrating superior performance consistency compared to other metaheuristic approaches. While all bio-inspired methods substantially outperform the baseline SVM (88.33%), the computational cost varies significantly, with GA-SVM offering the best time-accuracy trade-off (89.26% in 1,547.74 ss) and PSO-SVM requiring the highest processing time (3,722.63 s) for comparable accuracy. The results validate DOA's effectiveness in hyperparameter optimization for CLFP recognition, though the 588-fold increase in processing time compared to baseline SVM highlights the computational overhead inherent in population-based optimization approaches.