Improved Loss Function Research Articles

Improved Loss Function for Mass Segmentation in Mammography Images Using Density and Mass Size.

Mass segmentation is one of the fundamental tasks used when identifying breast cancer due to the comprehensive information it provides, including the location, size, and border of the masses. Despite significant improvement in the performance of the task, certain properties of the data, such as pixel class imbalance and the diverse appearance and sizes of masses, remain challenging. Recently, there has been a surge in articles proposing to address pixel class imbalance through the formulation of the loss function. While demonstrating an enhancement in performance, they mostly fail to address the problem comprehensively. In this paper, we propose a new perspective on the calculation of the loss that enables the binary segmentation loss to incorporate the sample-level information and region-level losses in a hybrid loss setting. We propose two variations of the loss to include mass size and density in the loss calculation. Also, we introduce a single loss variant using the idea of utilizing mass size and density to enhance focal loss. We tested the proposed method on benchmark datasets: CBIS-DDSM and INbreast. Our approach outperformed the baseline and state-of-the-art methods on both datasets.

Deep-computer-generated holography with temporal-focusing and a digital propagation matrix for rapid 3D multiphoton stimulation.

Deep learning-based computer-generated holography (DeepCGH) has the ability to generate three-dimensional multiphoton stimulation nearly 1,000 times faster than conventional CGH approaches such as the Gerchberg-Saxton (GS) iterative algorithm. However, existing DeepCGH methods cannot achieve axial confinement at the several-micron scale. Moreover, they suffer from an extended inference time as the number of stimulation locations at different depths (i.e., the number of input layers in the neural network) increases. Accordingly, this study proposes an unsupervised U-Net DeepCGH model enhanced with temporal focusing (TF), which currently achieves an axial resolution of around 5 µm. The proposed model employs a digital propagation matrix (DPM) in the data preprocessing stage, which enables stimulation at arbitrary depth locations and reduces the computation time by more than 35%. Through physical constraint learning using an improved loss function related to the TF excitation efficiency, the axial resolution and excitation intensity of the proposed TF-DeepCGH with DPM rival that of the optimal GS with TF method but with a greatly increased computational efficiency.

TU[formula omitted]Net-GAN: A temporal precipitation nowcasting model with multiple decoding modules

With the Earth’s temperature rising and abnormal weather events becoming frequent, the mechanisms of precipitation formation have become increasingly complex, leading to more significant spatiotemporal variability. This increased variability often results in severe flooding events. Despite extensive research on deep learning methods for rainfall prediction, challenges such as forecasting uncertainty and inaccurate predictions persist. Therefore, this study addresses these issues by proposing temporal U2Net (TU2Net) within the GAN framework. TU2Net comprises a nested UNet with two layers designed to handle time series data. The model also connects multiple Conv-GRU decoding modules to enhance spatiotemporal performance and improve prediction quality using an improved loss function. Experimental results demonstrate that TU2Net with the improved loss function outperforms DGMR and UNet in terms of prediction accuracy. Our project is open source and available on GitHub at https://github.com/clearlyzerolxd/TU2Net.

Computationally intelligent real-time security surveillance system in the education sector using deep learning.

Real-time security surveillance and identity matching using face detection and recognition are central research areas within computer vision. The classical facial detection techniques include Haar-like, MTCNN, AdaBoost, and others. These techniques employ template matching and geometric facial features for detecting faces, striving for a balance between detection time and accuracy. To address this issue, the current research presents an enhanced FaceNet network. The RetinaFace is employed to perform expeditious face detection and alignment. Subsequently, FaceNet, with an improved loss function is used to achieve face verification and recognition with high accuracy. The presented work involves a comparative evaluation of the proposed network framework against both traditional and deep learning techniques in terms of face detection and recognition performance. The experimental findings demonstrate that an enhanced FaceNet can successfully meet the real-time facial recognition requirements, and the accuracy of face recognition is 99.86% which fulfills the actual requirement. Consequently, the proposed solution holds significant potential for applications in face detection and recognition within the education sector for real-time security surveillance.

An Efficient Satellite Images Classification Approach Based on Fuzzy Cognitive Map Integration With Deep Learning Models Using Improved Loss Function

An Efficient Satellite Images Classification Approach Based on Fuzzy Cognitive Map Integration With Deep Learning Models Using Improved Loss Function

A multi‐class deep learning model to estimate brain age while addressing systematic bias of regression to the mean

AbstractBackgroundAge‐related changes in human brain may contribute to the development of age‐related neurodegenerative diseases. It may be possible to estimate “brain age” from magnetic resonance imaging (MRI) and the difference between a person’s brain and chronological age, ∆age, reflecting whether a person’s brain has been aging faster or slower than their chronological age. Deep Learning Models typically regress age on imaging features, leading to systematic biases associated with regression to the mean (RTM), including overestimation of brain age in younger persons and underestimation in older persons. We estimate brain age from a person’s MRI as a multi‐class classification problem and developed deep learning model to estimate brain age free from RTM bias.MethodTwo 3D ResNet‐18 models were implemented: regression and multi‐class classification. We transform the task of predicting age as continuous variable to predicting probabilities of discrete age values where age is discretized to closest integer values. Both models were trained on 7,372 T1‐weighted MRI scans of 5,848 cognitively normal participants (age: 8‐95 years) from public data sources (IXI, ICBM, ABIDE, NACC and OASIS). We create train, validation, and test set in ratio 80:10:10 with matching age distribution. Mean squared error is used as loss function in training regression model. Whereas in the classification model, we introduce two loss terms in addition to standard cross‐entropy loss: (1) minimizing the difference between mean (∑c∗pc) of expected age and the actual age, and (2) minimizing variance (∑(c‐mean)2∗pc), where pc is probability sample belonging to class c.ResultThe regression model achieved MAE = 3.93 years and R2 = 0.90 on unseen test set whereas classification model achieved MAE = 2.41 and R2 = 0.96 on same test set. We observe significant decrease in systematic bias using the classification model ‐ for younger (age<30) and older (age>70) subsets, average ∆age improved from 3.17 to 0.2, and from ‐2.49 to ‐0.97 respectively (Figure1).ConclusionOur proposed classification model with improved loss function to predict brain age from imaging features eliminates systemic bias present in traditional regression approaches and also improves performance by a significant margin. This model can be used more reliably to study age‐related alterations in brain and AD‐related deviations from natural aging.

Economy and carbon emissions optimization of different provinces or regions in China using an improved temporal attention mechanism based on gate recurrent unit

With the implementation of 14th Five-Year Plan in China and the completion of the poverty alleviation task, the economy in China has made great progress. However, the carbon dioxide (CO2) emission has also increased, which will lead to the greenhouse effect for threatening economic development and affecting people's lives. Global warming has become the greatest challenge to contemporary economic and social development, and the transition to low-carbon economy has become a general trend of world economic development. Meanwhile, the traditional neural network method for predicting the economy and CO2 emissions is not accurate and objective because of the unbalanced labeling of training data and the inability to capture the relationship between time series data. Therefore, an economy and CO2 emission prediction model based on temporal attention Gate Recurrent Unit (GRU) (TA-GRU) with loss function Balanced Mean Square Error (BMSE) (BTA-GRU) mechanism is proposed to analyze and optimize the energy structure of 27 provincial-level administrative regions in China. Due to the unbalanced data labels in the different energy structure of different regions, the improved loss function BMSE is used to solve the problem of unbalanced labels from the perspective of statistics. Then, with coal, gasoline, petroleum, coke, fuel oil, diesel, natural gas and crude oil as inputs, the per capita Gross Domestic Product (GDP) as the desired output and the CO2 as the undesired output, the GRU can better capture time step dependencies in time series, and the attention mechanism can assign different weights to each time step, so that the proposed method can better learn data features and improve the accuracy of the an economic and CO2 emission prediction model. Finally, the CO2 emission and economic prediction model of different provincial-level administrative regions in China is established based on this proposed method. The experimental results show that compared with other mainstream time series data prediction models, this proposed model has better stability and practicability, and its mean absolute error (MAE), mean absolute percentage error (MAPE) and root mean square error (RMSE) reach 3.43, 3.28% and 2449.57, respectively. Furthermore, according to the actual situation of efficient provincial administrative regions, the energy structure of inefficient provincial administrative regions can be adjusted and optimized. Such as the energy structure of Hebei Province can be adjusted according to the energy structure of Beijing, it is estimated that the CO2 emissions of Hebei will be reduced by 82,972.43 ×10000 tons (WT) and its per capita GDP will be increased by 115,856 Yuan. Moreover, Chinese provincial governments should vigorously promote the use of clean energy, accelerate the establishment of a sound green low-carbon circular development economic system, and help achieve the goal of carbon neutrality.

ClGanNet: A novel method for maize leaf disease identification using ClGan and deep CNN

With the advancement of technologies, automatic plant leaf disease detection has received considerable attention from researchers working in the area of precision agriculture. A number of deep learning-based methods have been introduced in the literature for automated plant disease detection. However, the majority of datasets collected from real fields have blurred background information, data imbalances, less generalization, and tiny lesion features, which may lead to over-fitting of the model. Moreover, the increased parameter size of deep learning models is also a concern, especially for agricultural applications due to limited resources. In this paper, a novel ClGan (Crop Leaf Gan) with improved loss function has been developed with a reduced number of parameters as compared to the existing state-of-the-art methods. The generator and discriminator of the developed ClGan have been encompassed with an encoder–decoder network to avoid the vanishing gradient problem, training instability, and non-convergence failure while preserving complex intricacies during synthetic image generation with significant lesion differentiation. The proposed improved loss function introduces a dynamic correction factor that stabilizes learning while perpetuating effective weight optimization. In addition, a novel plant leaf classification method ClGanNet, has been introduced to classify plant diseases efficiently. The efficiency of the proposed ClGan was validated on the maize leaf dataset in terms of the number of parameters and FID score, and the results are compared against five other state-of-the-art GAN models namely, DC-GAN, W-GAN, WGanGP, InfoGan, and LeafGan. Moreover, the performance of the proposed classifier, ClGanNet, was evaluated with seven state-of-the-art methods against eight parameters on the original, basic augmented, and ClGan augmented datasets. Experimental results of ClGanNet have outperformed all the considered methods with 99.97% training and 99.04% testing accuracy while using the least number of parameters.

Generative adversarial networks with multi-scale and attention mechanisms for underwater image enhancement

The images captured underwater are usually degraded due to the effects of light absorption and scattering. Degraded underwater images exhibit color distortion, low contrast, and blurred details, which in turn reduce the accuracy of marine biological monitoring and underwater object detection. To address this issue, a generative adversarial network with multi-scale and an attention mechanism is proposed to improve the quality of underwater images. To extract more effective features within the generative network, several modules are introduced: a multi-scale dilated convolution module, a novel attention module, and a residual module. These modules are utilized to design a generative network with a U-shaped structure. The multi-scale dilated convolution module is designed to extract features at multiple scales and expand the receptive field to capture more global information. The attention module directs the network’s focus towards important features, thereby reducing the interference from redundant feature information. To improve the discriminative power of the adversarial network, a multi-scale discriminator is designed. It has two output feature maps with different scales. Additionally, an improved loss function for the generative adversarial network is proposed. This improvement involves incorporating the total variation loss into the traditional loss function. The performance of different methods for enhancing underwater images is evaluated using the EUVP dataset and UIEB dataset. The experimental results demonstrate that the enhanced underwater images exhibit better quality and visual effects compared to other methods.

Application of FGD-BCEL loss function in segmenting temporal lobes on localized CT images for radiotherapy.

The aim of this study was to find a new loss function to automatically segment temporal lobes on localized CT images for radiotherapy with more accuracy and a solution to dealing with the classification of class-imbalanced samples in temporal lobe segmentation. Localized CT images for radiotherapy of 70 patients with nasopharyngeal carcinoma were selected. Radiation oncologists sketched mask maps. The dataset was randomly divided into the training set (n = 49), the validation set (n = 7), and the test set (n = 14). The training set was expanded by rotation, flipping, zooming, and shearing, and the models were evaluated using Dice similarity coefficient (DSC), Jaccard similarity coefficient (JSC), positive predictive value (PPV), sensitivity (SE), and Hausdorff distance (HD). This study presented an improved loss function, focal generalized Dice-binary cross-entropy loss (FGD-BCEL), and compared it with four other loss functions, Dice loss (DL), generalized Dice loss (GDL), Tversky loss (TL), and focal Tversky loss (FTL), using the U-Net model framework. With the U-Net model based on FGD-BCEL, the DSC, JSC, PPV, SE, and HD were 0.87 ± 0.11, 0.78 ± 0.11, 0.90 ± 0.10, 0.87 ± 0.13, and 4.11 ± 0.75, respectively. Except for the SE, all the other evaluation metric values of the temporal lobes segmented by the FGD-BCEL-based U-Net model were improved compared to the DL, GDL, TL, and FTL loss function-based U-Net models. Moreover, the FGD-BCEL-based U-Net model was morphologically more similar to the mask maps. The over- and under-segmentation was lessened, and it effectively segmented the tiny structures in the upper and lower poles of the temporal lobe with a limited number of samples. For the segmentation of the temporal lobe on localized CT images for radiotherapy, the U-Net model based on the FGD-BCEL can meet the basic clinical requirements and effectively reduce the over- and under-segmentation compared with the U-Net models based on the other four loss functions. However, there still exists some over- and under-segmentation in the results, and further improvement is needed.

End-to-end deep learning radiomics: development and validation of a novel attention-based aggregate convolutional neural network to distinguish breast diffuse large B-cell lymphoma from breast invasive ductal carcinoma.

Apart from invasive pathological examination, there is no effective method to differentiate breast diffuse large B-cell lymphoma (DLBCL) from breast invasive ductal carcinoma (IDC). In this study, we aimed to develop and validate an effective deep learning radiomics model to discriminate between DLBCL and IDC. A total of 324 breast nodules from 236 patients with baseline 18F-fluorodeoxyglucose (18F-FDG) positron emission tomography/computed tomography (PET/CT) were retrospectively analyzed. After grouping breast DLBCL and breast IDC patients, external and internal datasets were divided according to the data collected by different centers. Preprocessing was then used to process the original PET/CT images and an attention-based aggregate convolutional neural network (AACNN) model was designed. The AACNN model was trained using patches of CT or PET tumor images and optimized with an improved loss function. The final ensemble predictive model was built using distance weight voting. Finally, the model performance was evaluated and statistically verified. A total of 249 breast nodules from Fudan University Shanghai Cancer Center (FUSCC) and 75 breast nodules from Shanghai Proton and Heavy Ion Center (SPHIC) were selected as internal and external datasets, respectively. On the internal testing, our method yielded an area under the curve (AUC), accuracy (ACC), sensitivity (SEN), specificity (SPE), positive predictive value (PPV), negative predictive value (NPV), and harmonic mean of precision and sensitivity (F1) of 0.886, 83.0%, 80.9%, 85.0%, 84.8%, 81.2%, and 0.828, respectively. Meanwhile on the external testing, the results were 0.788, 71.6%, 61.4%, 84.7%, 84.0%, 62.6%, and 0.709, respectively. Our study outlines a deep learning radiomics method which can automatically, noninvasively, and accurately differentiate breast DLBCL from breast IDC, which will be more in line with the needs and strategies of precision medicine, individualized diagnosis, and treatment.

Remote Sensing Image Dehazing through an Unsupervised Generative Adversarial Network.

The degradation of visual quality in remote sensing images caused by haze presents significant challenges in interpreting and extracting essential information. To effectively mitigate the impact of haze on image quality, we propose an unsupervised generative adversarial network specifically designed for remote sensing image dehazing. This network includes two generators with identical structures and two discriminators with identical structures. One generator is focused on image dehazing, while the other generates images with added haze. The two discriminators are responsible for distinguishing whether an image is real or generated. The generator, employing an encoder-decoder architecture, is designed based on the proposed multi-scale feature-extraction modules and attention modules. The proposed multi-scale feature-extraction module, comprising three distinct branches, aims to extract features with varying receptive fields. Each branch comprises dilated convolutions and attention modules. The proposed attention module includes both channel and spatial attention components. It guides the feature-extraction network to emphasize haze and texture within the remote sensing image. For enhanced generator performance, a multi-scale discriminator is also designed with three branches. Furthermore, an improved loss function is introduced by incorporating color-constancy loss into the conventional loss framework. In comparison to state-of-the-art methods, the proposed approach achieves the highest peak signal-to-noise ratio and structural similarity index metrics. These results convincingly demonstrate the superior performance of the proposed method in effectively removing haze from remote sensing images.

Time Series Multi-Step Forecasting Based on Memory Network for the Prognostics and Health Management in Freight Train Braking System

In rail trains, the prognostics and health management system will make use of data from sensors deployed in key components. These data are typically non-stationary multi-variable time series with abrupt variations. The challenges of multivariate time series forecasting include how to identify the interactions between various variables, and how to assess the influence of historical data on present data. This paper presents the nature of the air brake system and its physical models. By analyzing the observed data obtained from the freight train air brake system, a Memory network-based Time series Multi-step fault Forecasting model (MTMF) is proposed. MTMF consists of a dual-encoder and a memory network module for feature extraction. The two encoders are used to extract features of short-term and long-term historical data respectively. The memory network is used to further mine the different weights of the long-term historical data. MTMF also designs a joint multi-step forecasting loss function, which is composed of shape, time and mean square error losses between the forecasting series and the real series. The improved loss function is no longer restricted to the differences in evaluation points. MTMF is evaluated on the real freight train braking system time series dataset. The results show that MTMF outperforms other forecasting methods on multivariate time series. The forecasting accuracy is increased by 5%-45%, demonstrating that the model is effective for non-stationary series and can better mine the dependence patterns for the train braking system.

SE-YOLOv7 Landslide Detection Algorithm Based on Attention Mechanism and Improved Loss Function

With the continuous development of computer vision technology, more and more landslide identification detection tasks have started to shift from manual visual interpretation to automatic computer identification, and automatic landslide detection methods based on remote sensing satellite images and deep learning have been gradually developed. However, most existing algorithms often have the problem of low precision and weak generalization in landslide detection. Based on the Google Earth Engine platform, this study selected landslide image data from 24 study areas in China and established the DN landslide sample dataset, which contains a total of 1440 landslide samples. The original YOLOv7 algorithm model was improved and optimized by applying the SE squeezed attention mechanism and VariFocal loss function to construct the SE-YOLOv7 model to realize the automatic detection of landslides in remote sensing images. The experimental results show that the mAP, Precision value, Recall value, and F1-Score of the improved SE-YOLOv7 model for landslide identification are 91.15%, 93.35%, 94.54%, and 93.94%, respectively. At the same time, through a field investigation and verification study in Qianyang County, Baoji City, Shaanxi Province, comparing the detection results of SE-YOLOv7, it is concluded that the improved SE-YOLOv7 can locate the landslide location more accurately, detect the landslide range more accurately, and have fewer missed detections. The research results show that the algorithm model has strong detection accuracy for many types of landslide image data, which provides a technical reference for future research on landslide detection based on remote sensing images.

Surface Defect Detection for Automated Tape Laying and Winding Based on Improved YOLOv5.

To address the issues of low detection accuracy, slow detection speed, high missed detection rate, and high false detection rate in the detection of surface defects on pre-impregnated composite materials during the automated tape laying and winding process, an improved YOLOv5 (You Only Look Once version 5) algorithm model was proposed to achieve the high-precision, real-time detection of surface defects. By leveraging this improvement, the necessity for frequent manual interventions, inspection interventions, and subsequent rework during the automated lay-up process of composite materials can be significantly reduced. Firstly, to improve the detection accuracy, an attention mechanism called "CA (coordinate attention)" was introduced to enhance the feature extraction ability, and a Separate CA structure was used to improve the detection speed. Secondly, we used an improved loss function "SIoU (SCYLLA-Intersection over Union) loss" to replace the original "CIoU (Complete-Intersection over Union) loss", which introduced an angle loss as a penalty term to consider the directional factor and improve the stability of the target box regression. Finally, Soft-SIoU-NMS was used to replace the original NMS (non-maximum suppression) of YOLOv5 to improve the detection of overlapping defects. The results showed that the improved model had a good detection performance for surface defects on pre-impregnated composite materials during the automated tape laying and winding process. The FPS (frames per second) increased from 66.7 to 72.1, and the mAP (mean average precision) of the test set increased from 92.6% to 97.2%. These improvements ensured that the detection accuracy, as measured by the mAP, surpassed 95%, while maintaining a detection speed of over 70 FPS, thereby meeting the requirements for real-time online detection.

Few-shot Object Detection Model based on Transfer Learning and Convolutional Neural Network

Aiming at the problem that Faster R-CNN fails to meet the current situation of few-shot object detection, an improved few-shot object detection model is proposed based on transfer learning and Faster R-CNN. Transfer learning algorithm is applied by transferring trained weights which from similar source dataset initially. The attention mechanism is added to the feature extraction module of the Faster R-CNN algorithm model and the idea of feature pyramid is introduced to fuse feature maps to build an effective feature extraction model. At the same time, a regularization term of background suppression is added to loss function of Faster R-CNN to improve the accuracy of object detection. The experimental results show that not only the feature extraction ability of the object detection network is strengthened, but also the improved loss function can effectively suppress the background interference and make the model training more adequate as a result. Experimental results show that the proposed network can converge faster and the detection accuracy is improved compared with other methods.

Deep Learning-Based Self-Admitted Technical Debt Detection Empirical Research

<p>Self-Admitted Technical Debt (SATD) is a workaround for current gains and subsequent software quality in software comments. Some studies have been conducted using NLP-based techniques or CNN-based classifiers. However, there exists a class imbalance problem in different software projects since the software code comments with SATD features are significantly less than those without Non-SATD. Therefore, to design a classification model with the ability of dealing with this class imbalance problem is necessary for SATD detection. We propose an improved loss function based on information entropy. Our proposed function is studied in a variety of application scenarios. Empirical research on 10 JAVA software projects is conducted to show the competitiveness of our new approach. We find our proposed approach can perform significantly better than state-of-the-art baselines. </p> <p> </p>

Using deep learning to track time × frequency whistle contours of toothed whales without human-annotated training data.

Many odontocetes produce whistles that feature characteristic contour shapes in spectrogram representations of their calls. Automatically extracting the time × frequency tracks of whistle contours has numerous subsequent applications, including species classification, identification, and density estimation. Deep-learning-based methods, which train models using analyst-annotated whistles, offer a promising way to reliably extract whistle contours. However, the application of such methods can be limited by the significant amount of time and labor required for analyst annotation. To overcome this challenge, a technique that learns from automatically generated pseudo-labels has been developed. These annotations are less accurate than those generated by human analysts but more cost-effective to generate. It is shown that standard training methods do not learn effective models from these pseudo-labels. An improved loss function designed to compensate for pseudo-label error that significantly increases whistle extraction performance is introduced. The experiments show that the developed technique performs well when trained with pseudo-labels generated by two different algorithms. Models trained with the generated pseudo-labels can extract whistles with an F1-score (the harmonic mean of precision and recall) of 86.31% and 87.2% for the two sets of pseudo-labels that are considered. This performance is competitive with a model trained with 12 539 expert-annotated whistles (F1-score of 87.47%).

Defects identification and location of underground space for ground penetrating radar based on deep learning

Collapses of urban roads caused by defects in shallow underground spaces are extremely dangerous, and this imposes high requirements for the detection of potential defect areas. Although the characteristics of defects, including voids, water-bearing voids, hyperbola defects and loose defects, can be effectively detected using ground penetrating radar (GPR), amounts of GPR data are currently limited. In this work, a deep convolutional generative adversarial network (DCGAN) with an improved loss function is applied to augment existing GPR data to give a total of 3,256 images. More importantly, a recognition model for underground defects is proposed based on GPR B-scans and strengthened with attention modules using YOLO v5. The mean average precision (mAP) of the model is 85.4 %, a value 3.26 % higher than that of YOLO v5. The values of the average precision (AP) for voids, water-bearing voids, hyperbolae, and loose defects are 87.5 %, 86.6 %, 96.2 %, and 71.1 %, respectively. Finally, the locations and span sizes of the defects are obtained by estimating the velocity of the electromagnetic wave, and this approach is verified through a field test. The absolute error in the burial depth is less than 0.16 m, and the average error ratio is less than 15 %. The absolute error in the horizontal span is found to be lower than 0.28 m, with an average error ratio of smaller than 22 %. Consequently, the proposed method provides reasonable support for more accurate and effective detection of underground defects.

A novel model with an improved loss function to predict the velocity field from the pressure on the surface of the hydrofoil

Reconstructing the flow field from limited observations is critical, as it can be used in marine applications which monitor flow changes with limited observations. With the development of deep learning, prediction models based on neural networks have been used for rapid prediction of flow fields. The work proposes models based on U-structure networks (U-net) and Signed Distance Function (SDF) to predict the velocity field from sparse pressure on the surface of the hydrofoil. An improved loss function is employed to improve the prediction accuracy. Through training and testing on a numerical simulation dataset, it is found that the prediction of models has a good agreement with the true values. The improved loss function redirects attention of the model to areas with poor prediction accuracy near the surface of the hydrofoil. For the dataset unknown to models, the prediction ability of the model is limited as input features are not enough to support strong generalization.