Cross-entropy Loss Function Research Articles

Enhancing cross entropy with a linearly adaptive loss function for optimized classification performance

We propose the linearly adaptive cross entropy loss function. This is a novel measure derived from the information theory. In comparison to the standard cross entropy loss function, the proposed one has an additional term that depends on the predicted probability of the true class. This feature serves to enhance the optimization process in classification tasks involving one-hot encoded class labels. The proposed one has been evaluated on a ResNet-based model using the CIFAR-100 dataset. Preliminary results show that the proposed one consistently outperforms the standard cross entropy loss function in terms of classification accuracy. Moreover, the proposed one maintains simplicity, achieving practically the same efficiency to the traditional cross entropy loss. These findings suggest that our approach could broaden the scope for future research into loss function design.

CNN-Based Distracted Driver Recognition: Enhancing Accuracy Through Optimized Training Techniques

Abstract. Distracted driving is a significant contributor to vehicle accidents worldwide, underscoring the need for an effective recognition model that can identify risky driver behaviors in real-time and provide timely alerts. This study proposes a Convolutional Neural Network (CNN)-based model designed to accurately detect distracted drivers. To optimize the model's performance, this paper implemented several strategic enhancements. During the preprocessing phase, 90% of the dataset was allocated for training, ensuring a comprehensive learning process, while the remaining 10% was used for validation to assess the model's accuracy. The Adam optimizer was chosen for its ability to dynamically adjust the learning rate, facilitating faster convergence to an optimal solution. Additionally, the Cross-Entropy loss function was employed to amplify errors during training, driving the model to correct inaccuracies more effectively. The model was trained over 25 epochs, resulting in an accuracy rate of nearly 80%. This level of performance demonstrates the models viability for real-world applications, where it can play a critical role in reducing accidents caused by distracted driving. Future research may focus on further refining the model by exploring advanced loss functions, optimizers, and CNN architectures, as well as incorporating more sophisticated data preprocessing techniques.

Transfer Learning Approaches for Brain Metastases Screenings

Background: In this study, we examined the effectiveness of transfer learning in improving automatic segmentation of brain metastases on magnetic resonance imaging scans, with potential applications in preventive exams and remote diagnostics. Methods: We trained three deep learning models on a public dataset from the ASNR-MICCAI Brain Metastasis Challenge 2024, fine-tuned them on a small private dataset, and compared their performance to models trained from scratch. Results: Results showed that models using transfer learning performed better than scratch-trained models, though the improvement was not statistically substantial. The custom Tversky and Binary Cross-Entropy loss function helped manage class imbalance and reduce false negatives, limiting missed tumor regions. Medical experts noted that, while fine-tuned models worked well with larger, well-defined tumors, they struggled with tiny, scattered tumors in complex cases. Conclusions: This study highlights the potential of transfer learning and tailored loss functions in medical imaging, while also pointing out the models’ limitations in detecting very small tumors in challenging cases.

AIpollen: An Analytic Website for Pollen Identification Through Convolutional Neural Networks

With the rapid development of artificial intelligence, deep learning has been widely applied to complex tasks such as computer vision and natural language processing, demonstrating its outstanding performance. This study aims to exploit the high precision and efficiency of deep learning to develop a system for the identification of pollen. To this end, we constructed a dataset across 36 distinct genera. In terms of model selection, we employed a pre-trained ResNet34 network and fine-tuned its architecture to suit our specific task. For the optimization algorithm, we opted for the Adam optimizer and utilized the cross-entropy loss function. Additionally, we implemented ELU activation function, data augmentation, learning rate decay, and early stopping strategies to enhance the training efficiency and generalization capability of the model. After training for 203 epochs, our model achieved an accuracy of 97.01% on the test set and 99.89% on the training set. Further evaluation metrics, such as an F1 score of 95.9%, indicate that the model exhibits good balance and robustness across all categories. To facilitate the use of the model, we develop a user-friendly web interface. Users can upload images of pollen grains through the URL link provided in this article) and immediately receive predicted results of their genus names. Altogether, this study has successfully trained and validated a high-precision pollen grain identification model, providing a powerful tool for the identification of pollen.

B-TBM: A Novel Deep Learning Model with Enhanced Loss Function for HAZOP Risk Classification Using Natural Language Statistical Laws

HAZOP is a paradigm of industrial safety, and the introduction of deep learning-based HAZOP text categorization marks the arrival of an intelligent era of safety analysis. However, existing risk analysis methods have limitations in processing complex texts and extracting deep risk features. To solve this problem, this paper proposes a novel HAZOP risk event classification model based on BERT, BiLSTM, and TextCNN. The complexity of HAZOP text is revealed by introducing statistical laws of natural language, such as Zipf’s law and Heaps’ law, and the outputs of different levels of BERT are further combined linearly to collaborate with BiLSTM and TextCNN to capture long-term dependency and local contextual information for a more accurate classification task. Meanwhile, an improved loss function is proposed to effectively solve the deficiencies of the traditional cross-entropy loss function in the mislabeling process and improve the generalization ability of the model. It is experimentally demonstrated that the accuracy of the model is improved by 3% to 4% compared to the traditional BERT model in the task of severity and possibility classification of HAZOP reports. This study not only improves the accuracy and efficiency of HAZOP risk analysis, but also provides new ideas and methods for the application of natural language processing in industrial safety.

Deep learning to combat knee osteoarthritis and severity assessment by using CNN-based classification

BackgroundIn today’s digital age, various diseases drastically reduce people’s quality of life. Arthritis is one amongst the most common and debilitating maladies. Osteoarthritis affects several joints, including the hands, knees, spine, and hips. This study focuses on the medical disorder underlying Knee Osteoarthritis (KOA) which severely impairs people’s quality of life. KOA is characterised by restricted mobility, stiffness, and terrible pain and can be caused by a range of factors such as ageing, obesity, and traumas. This degenerative disorder leads to progressive wear and tear of the knee joint.MethodsTo combat arthritis in the kneecap, this study employs a 12-layer Convolutional Neural Network (CNN) to reach deep learning capabilities. A collection of data from the Osteoarthritis Initiative (OAI) is used to classify KOA. Through the use of medical image processing; the study ascertains whether an individual has this ailment. A sophisticated CNN architecture created especially for binary classification and KOA severity utilising deep learning algorithms is the main component of this work.ResultsThe cross-entropy loss function is an important component of the model’s laborious design that classifies data into two groups. The remaining section uses the Kellgren-Lawrence (KL) grade to classify the disease’s severity. In the binary classification, the proposed algorithm outperforms previous methods with an accuracy rate of 92.3%, and in the multiclassification, its accuracy rate is 78.4% which is superior to the previous findings.ConclusionLooking ahead, the research broadens the scope of this work by gathering information from various sources and using these methods on a wider range of datasets and situations. The potential for major advancements in the field of osteoarthritis detection and classification is highlighted by this forward-looking approach. Furthermore, this method reduces the intervention of medical practitioners and ultimately results in accurate diagnosis.Clinical trial numberNot applicable.

Hybrid Mobile Robot Path Planning Using Safe JBS-A*B Algorithm and Improved DWA Based on Monocular Camera

This paper addresses the formidable challenge of enabling autonomous navigation in Mobile Robots (MRs), focusing on the development of advanced path planning strategies. Despite their pivotal role in diverse applications, they face challenges in dynamic settings due to limitation in existing Global Path Planning (GPP) and Local Path Planning (LPP) techniques. In response to this, we propose an innovative hybrid path planning approach that enhances the A* algorithm with a risk-aware heuristic function and integrates the Jump Point Search (JPS) technique for route optimization. Additionally, B-spline smoothing is employed for perceptually global trajectory refinement. Our approach also includes an innovative improvement to the Dynamic Window Approach (DWA) to align with the proposed enhanced A* algorithm for effective local navigation. Acknowledging the importance of high-quality input in path planning, we present substantial improvements to the IRDC-Net, a monocular-image semantic-segmentation model that we studied previously. Novel improvements include the integration of quantization and the Adam optimizer, along with the implementation of the Balanced Cross-Entropy loss function. These enhancements not only elevate the output quality of IRDC-Net but also reduce the model’s training parameters. The experimental results demonstrate the performance and viability of the proposed algorithm. Ultimately, the hybrid MR’s path planning algorithm exhibits proficiency across various tasks, particularly in addressing the challenge of evading moving obstacles to ensure the robot’s safety while adhering to the global path.

Image-Based Fitness Yoga Pose Recognition: Using Ensemble Learning and Multi-head Attention

With the increasing awareness of fitness, more and more people are choosing to participate in fitness activities. Yoga, as a form of exercise that improves both physical and mental health, is becoming increasingly popular worldwide. In order to assist yoga practitioners in more effective training through automated or semi automated systems, improve training effectiveness, assist professional athletes in training through intelligent recognition systems, correct movements, and improve athletic performance. This paper proposes a method that addresses the low accuracy issue of current yoga pose recognition algorithms by integrating multi-head attention mechanism and ensemble learning. Firstly, the Mixup algorithm is used to enhance yoga movement images. Subsequently, convolutional features are extracted from the images using the ResNet101 and VGGNet19 transfer learning models. Finally, the extracted convolutional features are combined and stacked using a multi-head attention mechanism. Model training, validation, and testing are performed using the Soft target cross-entropy loss function. Experimental results demonstrate that the proposed method achieves a training accuracy of 100%, a validation accuracy of 89.94%, a testing accuracy of 93.79%, and a detection speed of 297 frames per second. Overall, this method demonstrates high stability and robustness, providing a technological foundation for intelligent recognition of yoga poses.

Improving Effectiveness by Reducing Overconfidence in Large Catalogue Sequential Recommendation with gBCE loss

A large catalogue size is one of the central challenges in training recommendation models: a large number of items makes them memory and computationally inefficient to compute scores for all items during training, forcing these models to deploy negative sampling. However, negative sampling increases the proportion of positive interactions in the training data, and therefore, models trained with negative sampling tend to overestimate the probabilities of positive interactions – a phenomenon we call overconfidence . While the absolute values of the predicted scores/probabilities are not important for the ranking of retrieved recommendations, overconfident models may fail to estimate nuanced differences in the top-ranked items, resulting in degraded performance. In this paper, we show that overconfidence explains why the popular SASRec model underperforms when compared to BERT4Rec. This is contrary to the BERT4Rec authors’ explanation that the difference in performance is due to the bi-directional attention mechanism. To mitigate overconfidence, we propose a novel Generalised Binary Cross-Entropy Loss function (gBCE) and theoretically prove that it can mitigate overconfidence. We further propose the gSASRec model, an improvement over SASRec that deploys an increased number of negatives and the gBCE loss. Through detailed experiments on three datasets, we show that gSASRec does not exhibit the overconfidence problem. As a result, gSASRec can outperform BERT4Rec (e.g. +9.47% NDCG on the MovieLens-1M dataset), while requiring less training time (e.g. -73% training time on MovieLens-1M). Moreover, in contrast to BERT4Rec, gSASRec is suitable for large datasets that contain more than 1 million items. Finally, we show how addressing overconfidence can improve model calibration – the ability of a model to predict actual interaction probabilities accurately. By applying gBCE to the SASRec model on MovieLens-1M dataset, we reduce the models’ expected calibration error by 98.9% (from 0.966 to 0.01).

Unravelling sleep patterns: Supervised contrastive learning with self-attention for sleep stage classification

Sleep data scoring is a crucial and primary step for diagnosing sleep disorders to know the sleep stages from the PSG signals. This study uses supervised contrastive learning with a self-attention mechanism to classify sleep stages. We propose a deep learning framework for automatic sleep stage classification, which involves two training phases: (1) the feature representation learning phase, in which the feature representation network (encoder) learns to extract features from the electroencephalogram (EEG) signals, and (2) the classification network training phase, where a pre-trained encoder (trained during phase I) along with the classifier head is fine-tuned for the classification task. The PSG data shows a non-uniform distribution of sleep stages, with wake (W) (around 30% of total samples) and N2 stages (around 58% and 37% of total samples in Physionet EDF-Sleep 2013 and 2018 datasets, respectively) being more prevalent, leading to an imbalanced dataset. The imbalanced data issue is addressed using a weighted softmax cross-entropy loss function that assigns higher weights to minority sleep stages. Additionally, an oversampling technique (the synthetic minority oversampling technique (SMOTE) (Chawla et al., 2002)[1] ) is applied to generate synthetic samples for minority classes. The proposed model is evaluated on the Physionet EDF-Sleep 2013 and 2018 datasets using Fpz-Cz and Pz-Oz EEG channels. It achieved an overall accuracy of 94.1%, a macro F1 score of 92.64, and a Cohen’s Kappa coefficient of 0.92. Ablation studies demonstrated the importance of triplet loss-based pre-training and oversampling for enhancing performance. The proposed model requires minimal pre-processing, eliminating the need for extensive signal processing expertise, and thus is well-suited for clinicians diagnosing sleep disorders.

Explainable deep learning for glaucoma detection: Efficient dense net with quantum-based Dwarf Mongoose optimization

Glaucoma, a progressive eye disease, leads to irreversible vision loss due to optic nerve damage. This study aims to develop a more explainable and interpretable deep learning-based model for the detection and classification of glaucoma, addressing the current limitations in the field. The proposed approach utilizes a modified DenseNet-201 architecture, incorporating an additional transition layer for enhanced glaucoma classification. A pre-trained Efficient DenseNet is applied, and a reweighted cross-entropy loss function is used to handle the class imbalance in the training data. Quantum-based Dwarf Mongoose Optimization (QDMOA) is employed to fine-tune the model’s parameters, minimizing over fitting and improving generalization on small datasets. Experimental results demonstrate that the proposed method effectively detects and classifies two types of glaucoma. The use of dense connections with regularization significantly reduces over fitting, and the reweighted loss function improves model performance on imbalanced data. The developed deep learning model offers improved accuracy in glaucoma detection and classification while addressing the need for more interpretable and reliable models in medical applications. This research provides a practical tool for automated glaucoma screening, which could support ophthalmologists in early diagnosis and treatment, reducing the risk of irreversible vision loss.

HyFormer: a hybrid transformer-CNN architecture for retinal OCT image segmentation

Optical coherence tomography (OCT) has become the leading imaging technique in diagnosing and treatment planning for retinal diseases. Retinal OCT image segmentation involves extracting lesions and/or tissue structures to aid in the decisions of ophthalmologists, and multi-class segmentation is commonly needed. As the target regions often spread widely inside the retina, and the intensities and locations of different categories can be close, good segmentation networks must possess both global modeling capabilities and the ability to capture fine details. To address the challenge in capturing both global and local features simultaneously, we propose HyFormer, an efficient, lightweight, and robust hybrid network architecture. The proposed architecture features parallel Transformer and convolutional encoders for independent feature capture. A multi-scale gated attention block and a group positional embedding block are introduced within the Transformer encoder to enhance feature extraction. Feature integration is achieved in the decoder composed of the proposed three-path fusion modules. A class activation map-based cross-entropy loss function is also proposed to improve segmentation results. Evaluations are performed on a private dataset with myopic traction maculopathy lesions and the public AROI dataset for retinal layer and lesion segmentation with age-related degeneration. The results demonstrate HyFormer's superior segmentation performance and robustness compared to existing methods, showing promise for accurate and efficient OCT image segmentation. .

A sound-vibration physical-information fusion constraint-guided deep learning method for rolling bearing fault diagnosis

Although current deep learning models for bearing fault diagnosis have achieved excellent accuracy, the lack of constraint-guided learning of the physical mechanisms of real bearing failures and a physically scientific training paradigm leads to low interpretability and unreliability of intelligent fault diagnosis models. In this study, a sound-vibration physical-information fusion constraint-guided (PFCG) deep learning (DL) method is proposed, aiming at weighted fusion of sound and vibration multi-physical information into a deep learning model, to guide the DL model to learn more realistic physical laws of bearing failure. Firstly, a 15-degree-of-freedom nonlinear dynamics model of multi-stage degraded bearing failure mechanism with sound-vibration response is developed, which considers the evolutionary mechanism of bearing failure from healthy state to different stages, and utilizes a particle filtering algorithm for dynamic calibration of hidden parameters. Moreover, a lightweight DL fault diagnosis model is designed to realize the deep interaction between the physical model and the DL model through the weighted fusion of the cross-entropy loss function, physical consistency loss and uncertainty loss. Moreover, the superior diagnostic performance of the proposed sound and vibration PFCG-DL model is verified by comparing the performance fluctuations and parameter attributes of different DL benchmark models before and after being guided by physical information fusion constraints (PFCG). Eventually, the proposed PFCG-Transformer model achieves a diagnostic accuracy of 99.45% while keeping the number of parameters at only 0.62M, which significantly improves the accuracy and reduces the computational complexity by 81.5% compared to the CAME-Transformer model's 3.24 M number of parameters and 95.00% diagnostic accuracy. In addition, the test time of PFCG-Transformer is reduced to 1.02 s, which is 60.2% less than CAME-Transformer, demonstrating higher computational efficiency and real-time performance. Importantly, in terms of interpretability, the engineering interpretability and credibility of the models are further improved by visualizing the feature learning results of the vibration modal and multimodal fusion models and the sensitivity analyses of the sound-vibration response models with internal and external physical hyperparameters. Therefore, this study proposes a physical information-guided deep learning method with strong interpretability and superior performance, which provides an important reference for further research and application in the field of bearing fault diagnosis.

Detection of artificial spots in fundus images using modified U-Net based semantic segmentation

Today, deep learning algorithms are playing a very crucial role in the initial stage diagnosis of several fundus diseases like glaucoma, hypertension, and diabetic retinopathy. Lots of research is going on in this area now-a-days. During the acquisition of fundus images some artificial spots (e.g. because of the device itself, dust particles in the surroundings) have been added in captured images. In this paper, artificial spots in the fundus images that are generated due to the non-standardized conditions of scanning devices, are detected with the help of a newly proposed modified UNet (mU-Net) semantic segmentation model. Initially, preprocessing methods such as Gaussian blur, thresholding, and Hough transform have been used to create artificial spots. Now these preprocessed images have been used for training the proposed model. To make the proposed model more effective, the following modifications like, Regularization techniques (early stopping, greater weight decay, and Adam optimizer), decay learning rate scheduler, categorical cross-entropy loss function, and a significant number of filters have been modified in simple U-Net model. Apart from these mentioned modifications in the base U-Net model, a feature injecting module (FIM) has been added between the expansion and the contraction section of the simple U-Net model. FIM adds the features of the input image at the time of up-sampling. The addition of FIM to a simple U-Net model improves the detection of artificial spots and enhances the performance of the model. The mU-Net has been compared with other models, namely simple U-Net, V-Net, UNet++, ResUnet-a, WideU-Net, and Swin-Unet. The Friedman test that has been conducted on IOU, DICE, MAE, PSNR, and SSIM scores, found that mU-Net balances evaluation metrics well. It appears that the nonparametric Friedman test will improve reproducibility by demonstrating statistical significance. The IOU, DICE, MAE, PSNR, and SSIM scores of the proposed model indicate superior performance as compared to other models.

Abnormal recognition-assisted and onset-offset aware network for pathological wearable ECG delineation

Electrocardiogram (ECG) delineation is essential to the identification of abnormal cardiac status, especially when ECG signals are remotely monitored with wearable devices. The complexity and diversity of cardiac conditions generate numerous pathological ECG patterns, not only requiring the recognition of normal ECG but also addressing an extensive range of abnormal ECG patterns, posing a challenging task. Therefore, we propose an abnormal recognition-assisted network to integrate supplementary information on diverse ECG patterns. Simultaneously, we design an onset-offset aware loss to enhance precise waveform localization. Specifically, we establish a two-branch framework where ECG delineation serves as the target task, producing the final segmentation results. Additionally, the abnormal recognition-assisted network serves as an auxiliary task, extracting multi-label pathological information from ECGs. This joint learning approach establishes crucial correlations between ECG delineation and associated ECG abnormalities. The correlations enable the model to demonstrate sufficient generalization in the presence of diverse abnormal ECG patterns. Besides, onset-offset aware loss focuses intensively on wave onsets and offsets by applying biased weights to various waveform positions. This approach ensures a focus on precise localization, facilitating seamless integration into cross-entropy loss function. A large-scale wearable 12-lead dataset containing 4,913 signals is collected, offering an extensive range of ECG data for model training. Results demonstrate that our method achieves outstanding performance on two test datasets, attaining sensitivity of 94.97% and 94.27% and an error tolerance lower than 20 ms. Furthermore, our method is effective for various aberrant ECG signals, including ST-segment changes, atrial premature beats, and right and left bundle branch blocks.

PGNN-Net: Parallel Graph Neural Networks for Hyperspectral Image Classification Using Multiple Spatial-Spectral Features

Hyperspectral image (HSI) shows great potential for application in remote sensing due to its rich spectral information and fine spatial resolution. However, the high dimensionality, nonlinearity, and complex relationship between spectral and spatial features of HSI pose challenges to its accurate classification. Traditional convolutional neural network (CNN)-based methods suffer from detail loss in feature extraction; Transformer-based methods rely too much on the quantity and quality of HSI; and graph neural network (GNN)-based methods provide a new impetus for HSI classification by virtue of their excellent ability to handle irregular data. To address these challenges and take advantage of GNN, we propose a network of parallel GNNs called PGNN-Net. The network first extracts the key spatial-spectral features of HSI using principal component analysis, followed by preprocessing to obtain two primary features and a normalized adjacency matrix. Then, a parallel architecture is constructed using improved GCN and ChebNet to extract local and global spatial-spectral features, respectively. Finally, the discriminative features obtained through the fusion strategy are input into the classifier to obtain the classification results. In addition, to alleviate the over-fitting problem, the label smoothing technique is embedded in the cross-entropy loss function. The experimental results show that the average overall accuracy obtained by our method on Indian Pines, Kennedy Space Center, Pavia University Scene, and Botswana reaches 97.35%, 99.40%, 99.64%, and 98.46%, respectively, which are better compared to some state-of-the-art methods.

Method for Augmenting Side-Scan Sonar Seafloor Sediment Image Dataset Based on BCEL1-CBAM-INGAN.

In this paper, a method for augmenting samples of side-scan sonar seafloor sediment images based on CBAM-BCEL1-INGAN is proposed, aiming to address the difficulties in acquiring and labeling datasets, as well as the insufficient diversity and quantity of data samples. Firstly, a Convolutional Block Attention Module (CBAM) is integrated into the residual blocks of the INGAN generator to enhance the learning of specific attributes and improve the quality of the generated images. Secondly, a BCEL1 loss function (combining binary cross-entropy and L1 loss functions) is introduced into the discriminator, enabling it to focus on both global image consistency and finer distinctions for better generation results. Finally, augmented samples are input into an AlexNet classifier to verify their authenticity. Experimental results demonstrate the excellent performance of the method in generating images of coarse sand, gravel, and bedrock, as evidenced by significant improvements in the Frechet Inception Distance (FID) and Inception Score (IS). The introduction of the CBAM and BCEL1 loss function notably enhances the quality and details of the generated images. Moreover, classification experiments using the AlexNet classifier show an increase in the recognition rate from 90.5% using only INGAN-generated images of bedrock to 97.3% using images augmented using our method, marking a 6.8% improvement. Additionally, the classification accuracy of bedrock-type matrices is improved by 5.2% when images enhanced using the method presented in this paper are added to the training set, which is 2.7% higher than that of the simple method amplification. This validates the effectiveness of our method in the task of generating seafloor sediment images, partially alleviating the scarcity of side-scan sonar seafloor sediment image data.

A parallel neural networks for emotion recognition based on EEG signals

Our study proposes a novel Parallel Temporal–Spatial-Frequency Neural Network (PTSFNN) for emotion recognition. The network processes EEG signals in the time, frequency, and spatial domains simultaneously to extract discriminative features. Despite its relatively simple architecture, the proposed model achieves superior performance. Specifically, PTSFNN first applies wavelet transform to the raw EEG signals and then reconstructs the coefficients based on frequency hierarchy, thereby achieving frequency decomposition. Subsequently, the core part of the network performs three independent parallel convolution operations on the decomposed signals, including a novel graph convolutional network. Finally, an attention mechanism-based post-processing operation is designed to effectively enhance feature representation. The features obtained from the three modules are concatenated for classification, with the cross-entropy loss function being adopted. To evaluate the model’s performance, extensive experiments are conducted on the SEED and SEED-IV public datasets. The experimental results demonstrate that PTSFNN achieves excellent performance in emotion recognition tasks, with classification accuracies of 87.63% and 74.96%, respectively. Comparative experiments with previous state-of-the-art methods confirm the superiority of our proposed model, which can efficiently extract emotion information from EEG signals.

Dense pedestrian face detection in complex environments

To address the problem of dense crowd face detection in complex environments, this paper proposes a face detection model named Deep and Compact Face Detection (DCFD), which adopts an improved lightweight EfficientNetV2 network to replace the backbone network of RetinaFace. A large kernel attention mechanism is introduced to address the face detection task more accurately. The backbone network, an improved efficient channel attention (ECA) mechanism, is added to further improve the algorithm performance. The feature fusion module is an improved neural architecture search feature pyramid network (NAS-FPN) that significantly improves the face detection accuracy in different scenes. To balance the training process of positive and negative samples, we use the focus loss function to replace the traditional cross-entropy loss function. In different environments, the DCFD algorithm has shown efficient face detection performance. This algorithm provides not only a feasible and effective solution for solving the problem of face detection in dense groups but also an important basis for improving the accuracy of face detection models in practical applications.

White blood cell classification network using MobileNetv2 with multiscale feature extraction module and attention mechanism

White blood cells play a crucial role in the human immune system. The accurate classification of white blood cells can help doctors diagnose various diseases for patients. To enhance the classification accuracy of white blood cells micro-vision images, an efficient lightweight deep learning network called ICAFF-MobileNetv2 is proposed in this paper. Firstly, the pruning operations are applied to the MobileNetv2 and the multiscale feature extraction module is proposed to reduce the model size while ensuring the classification accuracy. Secondly, the improved coordinate attention mechanism module and the attention feature fusion module are incorporated into the inverted residual structures to enhance the feature extraction capability and the classification accuracy of the network. Furthermore, considering the imbalance among different types of white blood cells in the datasets, the method of label smoothing is introduced to improve the cross-entropy loss function and further enhance the classification accuracy. Finally, the improved network is validated on the proposed mixed dataset. The experimental results demonstrate that the proposed network features high classification accuracy and small model size. Compared to other classification networks, ICAFF-MobileNetv2 exhibits superior performance with 98.54 %, 98.21 %, 98.56 %, and 98.38 % of the classification accuracy, recall, precision and F1 score respectively, which are 0.7 %, 0.86 %, 0.92 %, and 0.89 % better than MobileNetv2.