Ablation Experiments Research Articles

VM-YOLO: YOLO with VMamba for Strawberry Flowers Detection

Computer vision technology is widely used in smart agriculture, primarily because of its non-invasive nature, which avoids causing damage to delicate crops. Nevertheless, the deployment of computer vision algorithms on agricultural machinery with limited computing resources represents a significant challenge. Algorithm optimization with the aim of achieving an equilibrium between accuracy and computational power represents a pivotal research topic and is the core focus of our work. In this paper, we put forward a lightweight hybrid network, named VM-YOLO, for the purpose of detecting strawberry flowers. Firstly, a multi-branch architecture-based fast convolutional sampling module, designated as Light C2f, is proposed to replace the C2f module in the backbone of YOLOv8, in order to enhance the network’s capacity to perceive multi-scale features. Secondly, a state space model-based lightweight neck with a global sensitivity field, designated as VMambaNeck, is proposed to replace the original neck of YOLOv8. After the training and testing of the improved algorithm on a self-constructed strawberry flower dataset, a series of experiments is conducted to evaluate the performance of the model, including ablation experiments, multi-dataset comparative experiments, and comparative experiments against state-of-the-art algorithms. The results show that the VM-YOLO network exhibits superior performance in object detection tasks across diverse datasets compared to the baseline. Furthermore, the results also demonstrate that VM-YOLO has better performances in the mAP, inference speed, and the number of parameters compared to the YOLOv6, Faster R-CNN, FCOS, and RetinaNet.

A Fault Diagnosis Model for Insulator Instance Segmentation in Transmission Lines Based on ELD-YOLOv8

Abstract A novel ELD-YOLOv8n transmission line insulator instance segmentation fault diagnosis model is proposed to accurately and meticulously segment every fault of the transmission line insulator. Firstly, an innovative and efficient lightweight downsampling module (ELD) was proposed, Efficient-Lightweight downsampling, This module is used to replace the standard downsampling unit in the model, which not only reduces the number of model parameters, but also enhances the feature extraction ability of the model; Then, a lightweight CARAFE module was used to replace the upsampling of the model, optimizing the upsampling process and reducing the number of parameters; Finally, CGAFsion is used to fuse the features extracted from the backbone network with the head features, effectively compensating for the information loss caused by the convolution process. The experimental results show that the improved model proposed in this study mAP@50 The indicator reached 86.2%. The effectiveness of the im-provement and significant instance segmentation fault detection capability have been demonstrated through ablation experiments. This study provides a new technical path for fault diagnosis of insulator instance segmentation in transmission lines.

Multi-Step Peak Passenger Flow Prediction of Urban Rail Transit Based on Multi-Station Spatio-Temporal Feature Fusion Model

Accurate prediction of station passenger flow is crucial for optimizing rail transit efficiency, but peak passenger flow in urban rail transit (URT) is often disrupted by random events, making predictions challenging. In this paper, in order to solve this challenge, the Bi-graph Graph Convolutional Spatio-Temporal Feature Fusion Network (BGCSTFFN)-based model is introduced to capture complex spatio-temporal correlations. A combination of a graph convolutional neural network and a Transformer is used. The model separately inputs land use (point of interest, POI) and station adjacency information as features into the BGCSTFFN model, using the Pearson correlation coefficient matrix, which is evaluated on real passenger flow dataset from 1 to 25 January 2019 in Hangzhou. The results showed that the model consistently provided the best prediction results across different datasets and prediction tasks compared to other baseline models. In addition, in tasks involving predictions with different combinations of inputs and prediction steps, the model showed superior performance at multiple prediction steps. Its practical application is validated by comparing the results of passenger flow prediction for different types of stations. In addition, the impact of these features on the prediction accuracy and the generalization ability of the model were verified by designing ablation experiments and testing on different datasets.

Identification of the number of leaks in water supply pipes based on wavelet scattering network and Bi-LSTM model with Bayesian optimization

Water supply pipeline leaks have significant impacts on urban areas and daily life. Acoustic techniques are widely applied in its leak detection, but there is a lack of research on quantifying multiple leaks in pipelines. This study presents a novel approach using a Wavelet Scattering Network (WSN) and Bi-directional Long and Short-Term Memory (Bi-LSTM) model with Bayesian optimization to identify leak quantities accurately. The WSN autonomously extracts signal features, which are subsequently processed by the Bi-LSTM for classification. Through Bayesian optimization, the model’s hyperparameters are refined to enhance performance. The results exhibit an impressive 90.8 % classification accuracy, showcasing the model’s effectiveness in accurately identifying leak quantities. In comparison with existing models, the proposed method stands out for its superior recognition accuracy, shorter training time, and enhanced noise immunity. Ablation experiments further confirm the distinct roles and contributions of the WSN, Bayesian optimization, and Bi-LSTM components in achieving these exceptional results.

A novel intelligent health indicator using acoustic waves: CEEMDAN-driven semi-supervised ensemble deep learning

Designing health indicators (HIs) for aerospace composite structures that demonstrate their health comprehensively, including all types of damage that can be adaptively updated, is challenging, especially under complex conditions like impact and compression-fatigue loadings. This paper introduces a new AI-based approach to designing reliable HIs (fulfilling requirements—monotonicity, prognosability, and trendability—referred to as ’Fitness’) for single-stiffener composite panels under fatigue loading using acoustic emission sensors. It incorporates complete ensemble empirical mode decomposition with adaptive noise for feature extraction, semi-supervised base deep learner models made of long short-term memory layers for information fusion, and a semi-supervised paradigm to simulate labels inspired by the physics of progressive damage. In this way, nondifferentiable prognostic criteria are implicitly implemented into the learning process. Ensemble learning, especially using a semi-supervised network built with bidirectional long short-term memory, improves HI quality while reducing deep learning randomness. The Fitness function equation has been modified to provide a more trustworthy foundation for comparison and enhance the practical reliability of the standard in prognostics and health management. Ablation experiments are conducted, including variations in dataset division and leave-one-out cross-validation, confirming the generalizability of the approach.

Research on failure diagnosis analysis of plunger gas lift system using convolutional neural network with multi-scale channel attention mechanism based on wavelet transform

Plunger gas lift technology has been extensively utilized in unconventional gas fields, owing to its distinct engineering benefits. However, as development challenges intensify, fault diagnosis has grown more intricate, and conventional diagnostic techniques exhibit delays and inaccuracies. With advancements in computer science, machine learning methods have demonstrated their prowess in establishing robust correlations between data features and prediction outcomes. Consequently, this paper introduces a convolutional neural network fault diagnosis model that incorporates a multi-scale channel attention mechanism based on wavelet transform. This model dissects features across various scales using wavelet transform and leverages channel attention to adaptively select channels encompassing fault features, thereby enhancing diagnostic and recognition accuracy. Furthermore, the model integrates a hyperparameter search optimization algorithm to refine the model’s architecture and comprehensively bolster its generalization capability. A comparison with actual field data reveals that the fault diagnosis accuracy of the WT-MACNN model stands at 83.33%. Digital ablation experiments underscore the limited accuracy of the basic CNN model in fault diagnosis, but the sequential introduction of the channel attention mechanism and wavelet transform layers significantly elevates model performance. When juxtaposed with SVM, KNN, and decision tree models, the WT-MACNN model exhibits a diagnostic accuracy improvement of 73.81%, 57.13%, and 54.76%, respectively. Additionally, to assess the model’s adaptability under diverse well conditions, this study reacquired 128 sets of field data. The verification results indicate that the model’s prediction accuracy across different well conditions is 83.59%. Despite occasional misjudgments among certain operating conditions, the overall performance remains outstanding, offering dependable support for diagnosing real-world scenarios. The outcomes of numerical experiments highlight the profound advantages of the proposed deep learning model in terms of generalization ability and diagnostic accuracy, validating its superiority in plunger gas lift fault identification and carrying substantial guidance for the application of this technology.

Research on Blood Cell Image Detection Method Based on Fourier Ptychographic Microscopy

Autonomous Fourier Ptychographic Microscopy (FPM) is a technology widely used in the field of pathology. It is compatible with high resolution and large field-of-view imaging and can observe more image details. Red blood cells play an indispensable role in assessing the oxygen-carrying capacity of the human body and in screening for clinical diagnosis and treatment needs. In this paper, the blood cell data set is constructed based on the FPM system experimental platform. Before training, four enhancement strategies are adopted for the blood cell image data to improve the generalization and robustness of the model. A blood cell detection algorithm based on SCD-YOLOv7 is proposed. Firstly, the C-MP (Convolutional Max Pooling) module and DELAN (Deep Efficient Learning Automotive Network) module are used in the feature extraction network to optimize the feature extraction process and improve the extraction ability of overlapping cell features by considering the characteristics of channels and spatial dimensions. Secondly, through the Sim-Head detection head, the global information of the deep feature map (mean average precision) and the local details of the shallow feature map are fully utilized to improve the performance of the algorithm for small target detection. MAP is a comprehensive indicator for evaluating the performance of object detection algorithms, which measures the accuracy and robustness of a model by calculating the average precision (AP) under different categories or thresholds. Finally, the Focal-EIoU (Focal Extended Intersection over Union) loss function is introduced, which not only improves the convergence speed of the model but also significantly improves the accuracy of blood cell detection. Through quantitative and qualitative analysis of ablation experiments and comparative experimental results, the detection accuracy of the SCD-YOLOv7 algorithm on the blood cell data set reached 92.4%, increased by 7.2%, and the calculation amount was reduced by 14.6 G.

Gesture recognition from surface electromyography signals based on the SE-DenseNet network.

In recent years, significant progress has been made in the research of gesture recognition using surface electromyography (sEMG) signals based on machine learning and deep learning techniques. The main motivation for sEMG gesture recognition research is to provide more natural, convenient, and personalized human-computer interaction, which makes research in this field have considerable application prospects in rehabilitation technology. However, the existing gesture recognition algorithms still need to be further improved in terms of global feature capture, model computational complexity, and generalizability. This paper proposes a fusion model of Squeeze-and-Excitation Networks (SE) and DenseNet, inserting attention mechanism between DenseBlock and Transition to focus on the most important information, improving feature representation ability, and effectively solving the problem of gradient vanishing. This proposed method was tested on the electromyographic gesture datasets NinaPro DB2 and DB4, achieving accuracies of 85.93 and 82.39 % respectively. Through ablation experiments, it was found that the method based on DenseNet-101 as the backbone model produced the best results. Compared with existing models, this proposed method has better robustness and generalizability in gesture recognition, providing new ideas for the development of sEMG signal gesture recognition applications in the future.

Semi-Supervised Chinese Word Segmentation in Geological Domain Using Pseudo-Lexicon and Self-Training Strategy

Chinese word segmentation (CWS), which involves splitting the sequence of Chinese characters into words, is a key task in natural language processing (NLP) for Chinese. However, the complexity and flexibility of geologic terms require that domain-specific knowledge be utilized in CWS for geoscience domains. Previous studies have identified several challenges that have an impact on CWS in the geoscience domain, including the absence of abundant labeled data and difficult-to-delineate complex geological word boundaries. To solve these problems, a novel semi-supervised deep learning framework, GeoCWS, is developed for CWS in the geoscience domain. The framework is designed with domain-enhanced features and an uncertainty-aware self-training strategy. First, n-grams are automatically constructed from the input text as a pseudo-lexicon. Then, a backbone model is suggested that learns domain-enhanced features by introducing a pseudo-lexicon-based memory mechanism to delineate complex geological word boundaries based on BERT. Next, the backbone model is fine-tuned with a small amount of labeled data to obtain the teacher model. Finally, we design a self-training strategy with joint confidence and uncertainty awareness to improve the generalization ability of the backbone model to unlabeled data. Our method outperformed the state-of-the-art baseline methods in extensive experiments, and ablation experiments verified the effectiveness of the proposed backbone model and self-training strategy.

Segmentation of coronary artery and calcification using prior knowledge based deep learning framework.

Computed tomography angiography (CTA) is used to screen for coronary artery calcification. As the coronary artery has complicated structure and tiny lumen, manual screening is a time-consuming task. Recently, many deep learning methods have been proposed for the segmentation (SEG) of coronary artery and calcification, however, they often neglect leveraging related anatomical prior knowledge, resulting in low accuracy andinstability. This study aims to build a deep learning based SEG framework, which leverages anatomical prior knowledge of coronary artery and calcification, to improve the SEG accuracy. Moreover, based on the SEG results, this study also try to reveal the predictive ability of the volume ratio of coronary artery and calcification for rotational atherectomy (RA). We present a new SEG framework, which is composed of four modules: the variational autoencoder based centerline extraction (CE) module, the self-attention (SA) module, the logic operation (LO) module, and the SEG module. Specifically, the CE module is used to crop a series of 3D CTA patches along the coronary artery, from which the continuous property of vessels can be utilized by the SA module to produce vessel-related features. According to the spatial relations between coronary artery lumen and calcification regions, the LO module with logic union and intersection is designed to refine the vessel-related features into lumen- and calcification-related features, based on which SEG results can be produced by the SEGmodule. Experimental results demonstrate that our framework outperforms the state-of-the-art methods on CTA image dataset of 72 patients with statistical significance. Ablation experiments confirm that the proposed modules have positive impacts to the SEG results. Moreover, based on the volume ratio of segmented coronary artery and calcification, the prediction accuracy of RA is 0.75. Integrating anatomical prior knowledge of coronary artery and calcification into the deep learning based SEG framework can effectively enhance the performance. Moreover, the volume ratio of segmented coronary artery and calcification is a good predictive factor forRA.

Adaptive Multi-Kernel Graph Neural Network for Drug-Drug Interaction Prediction.

Combination therapy, which synergistically enhances treatment efficacy and inhibits disease progression through the combined effects of multiple drugs, has emerged as a mainstream approach for treating complex diseases and alleviating symptoms. However, drug-drug interactions (DDIs) can sometimes lead to adverse reactions, potentially endangering lives. Therefore, developing efficient and accurate DDI prediction methods is crucial for elucidating drug mechanisms and preventing side effects. Current prediction methods often focus solely on the presence of interactions between drugs when constructing DDI graphs, neglecting the specific types of DDIs. This oversight can result in a decline in predictive performance. To address this issue, we propose an Adaptive Multi-Kernel Graph Neural Network (AMKGNN) for DDI prediction. AMKGNN differentiates DDIs into increase-type and decrease-type interactions, constructing separate increased DDI and decreased DDI graphs as convolutional kernels. AMKGNN employs a graph kernel learning mechanism that adaptively determines the optimal threshold between high-frequency and low-frequency signals in the network to capture node embeddings. Initially, AMKGNN learns drug embedding representations based on these two graph convolutional kernels and various drug features. These representations are then concatenated and input into a deep neural network to predict potential DDIs. The results show that our model achieved AUC and AUPR values above 90% across three sub-tasks on two datasets, significantly outperforming the other five comparison models. Furthermore, ablation experiments and case studies validate the superiority of AMKGNN.

A Flat-Span Contrastive Learning Method for Nested Named Entity Recognition

In Natural Language Processing (NLP), it is common for one entity to contain another entity, i.e., nested entities. However, the most commonly used methods can only handle flat entities but not nested entities. To solve this problem, this paper proposes a flat-span contrastive learning method for nested Named Entity Recognition (NER), which consists of two sub-modules: a flat NER module and a candidate span classification module. The flat NER module is used to recognize the outermost entities, and we use star-transformer to capture the long-range dependencies of sentences, and the Conditional Random Field (CRF) to decode the outermost entity spans, contrastive learning is introduced, and the InfoNEC loss function is used to increase the difference between entity spans and nonentity spans. Finally, to improve the model performance and reduce error propagation, we jointly train the flat NER and candidate span classification modules through multi-task learning. Experimental results on the GENIA, GermEval2014, and JNLPBA datasets thoroughly verify the effectiveness of our model, and the ablation experiments further demonstrate the effectiveness of the model components. In the candidate span classification module, we generate all possible candidate spans based on the outermost entities by enumeration to better distinguish entity spans from nonentity.

Hyperspectral Band Selection Method Based on Global Partition Clustering

Band selection is an important step in the dimensionality reduction processing of hyperspectral images and is highly important for eliminating redundant spectral information and reducing computational costs. In recent years, band selection methods based on ordered partition have been widely used in the dimensionality reduction processing of hyperspectral images. However, most of the methods use coarse and fine partition for band subspace partition, so that the partition results are affected by equal interval partition. Furthermore, existing methods usually select a representative band in each band subspace but do not consider the relationship between the output bands in the selection process, resulting in a certain degree of redundancy between the output bands. To solve the above problems, we propose a band selection method based on global partition clustering that contains band subspace partition and band selection. The band subspace partition is based on coarse and fine partition and the similarity-based ranking-structural similarity method, which divides the band space into band subspaces according to the relationship between the number of bands in the hyperspectral image and the number of selected bands. Band selection is based on the sequential forward selection method, which iteratively selects one band as the output band in each band subspace. The proposed method makes two main contributions. Firstly, this method avoids the negative effect of equal interval partition on the results. Secondly, this method fully considers the relationship between the selected bands and the bands to be selected. The effectiveness of the proposed method is demonstrated via ablation and comparison experiments on three publicly available datasets. Comparison experiments show that the classification accuracy of the proposed method can exceed the accuracy of the comparison methods.

Subgraph Topology and Dynamic Graph Topology Enhanced Graph Learning and Pairwise Feature Context Relationship Integration for Predicting Disease-Related miRNAs.

As an increasing number of microRNAs (miRNAs) have become biomarkers of various human diseases, prediction of the candidate disease-related miRNAs is helpful for facilitating the early diagnosis of diseases. Most of the recent prediction models concentrated on learning of the features from the heterogeneous graph composed of miRNAs and diseases. However, they failed to fully exploit the subgraph structures consisting of multiple miRNA and disease nodes, and they also did not completely integrate the context relationships among the pairwise features. We proposed a prediction model, SFPred, to integrate and encode the local topologies from neighborhood subgraphs, the dynamically evolved heterogeneous graph topology, and the context among pairwise features. First, the importance of an miRNA (disease) node to another node is formulated according to the subgraphs composed of their neighbors. Second, the features of each miRNA (disease) node continuously change when the graph encoding gradually deepens for the miRNA-disease heterogeneous network. A strategy based on multi-layer perceptron (MLP) is designed to estimate the edge weights according to the changed node features and form the dynamic graph topology. Third, considering the context relationships among the features of a pair of miRNA and disease nodes, a context relationship sensitive transformer is constructed to integrate these relationships. Finally, since the previous encoding layer of the transformer contains more detailed features of the pairwise, we present a multiperspective residual strategy to supplement the detailed features to the following encoding layer from the channel perspective and the feature one, respectively. The extensive experiments confirmed that SFPred outperforms eight state-of-the-art methods for the prediction of miRNA-disease associations, and the ablation experiments validate the effectiveness of the proposed innovations. The recall rates for the top-ranked candidate miRNAs related to the diseases and the case studies on three diseases indicate SFPred's ability in screening the reliable candidates for subsequent biological experiments.

Aspect category sentiment analysis based on pre-trained BiLSTM and syntax-aware graph attention network

Aspect Category Sentiment Analysis (ACSA) is a fine-grained sentiment analysis task aimed at predicting the sentiment polarity associated with aspect categories within a sentence.Most existing ACSA methods are based on a given aspect category to locate sentiment words related to it. When irrelevant sentiment words have semantic meaning for the given aspect category, it may cause the problem that sentiment words cannot be matched with aspect categories. To address the aforementioned issue, this paper proposes a novel approach for ACSA utilizing pre-trained Bidirectional Long Short-Term Memory (BiLSTM) and syntax-aware graph attention network. To address the issue of insufficient existing annotated datasets, a method of using transfer learning is proposed. Firstly, the BiLSTM model is used to pre-train on the document-level sentiment analysis dataset, and the obtained pre-training parameters are transferred to the aspect-level task model. Then, a syntax-aware graph attention network model is proposed to make full use of the syntactic structure and semantic information in the text, and combine the knowledge learned in pre-training to achieve the ACSA task. The performance evaluation of this method is carried out on five user comment text datasets, and the comprehensive ablation experiments prove that this method performs best compared with baseline models.

Herb-disease association prediction model based on network consistency projection

A growing number of biological and clinical reports indicate the usefulness of herbs in the treatment of complex human diseases, giving an essential supplement for modern medicine. Similar to drugs, the use of experimental validation to identify related diseases of given herbs is both expensive and time-consuming. Such validation is even more difficult because each herb always contains several components. It is alternative to design computational models to predict herb-disease associations (HDAs). Nevertheless, only a few computational models have been developed for HDA prediction. In this study, we make full use of several properties of herbs and diseases, which are collected in a public database HERB, to design a model named HDAPM-NCP for predicting HDAs. Based on these properties, six herb kernels and five disease kernels are constructed, which are further fused into one unified herb kernel and one disease kernel. These kernels and herb-disease adjacency matrix are fed into network consistency projection to quantify the strength of herb-disease pairs. The cross-validation results show the high performance of HDAPM-NCP. Such performance is higher than that of two previous models. The ablation experiments prove the effects of modules in this model. Finally, we also analyze the weakness and strength of the model, uncovering which herb-disease pairs that HDAPM-NCP can yield reliable or unsatisfied predictions, and a case study is conducted to prove that HDAPM-NCP can discover latent HDAs.

RST-DeepLabv3+: Multi-Scale Attention for Tailings Pond Identification with DeepLab

Tailing ponds are used to store tailings or industrial waste discharged after beneficiation. Identifying these ponds in advance can help prevent pollution incidents and reduce their harmful impacts on ecosystems. Tailing ponds are traditionally identified via manual inspection, which is time-consuming and labor-intensive. Therefore, tailing pond identification based on computer vision is of practical significance for environmental protection and safety. In the context of identifying tailings ponds in remote sensing, a significant challenge arises due to high-resolution images, which capture extensive feature details—such as shape, location, and texture—complicated by the mixing of tailings with other waste materials. This results in substantial intra-class variance and limited inter-class variance, making accurate recognition more difficult. Therefore, to monitor tailing ponds, this study utilized an improved version of DeepLabv3+, which is a widely recognized deep learning model for semantic segmentation. We introduced the multi-scale attention modules, ResNeSt and SENet, into the DeepLabv3+ encoder. The split-attention module in ResNeSt captures multi-scale information when processing multiple sets of feature maps, while the SENet module focuses on channel attention, improving the model’s ability to distinguish tailings ponds from other materials in images. Additionally, the tailing pond semantic segmentation dataset NX-TPSet was established based on the Gauge-Fractional-6 image. The ablation experiments show that the recognition accuracy (intersection and integration ratio, IOU) of the RST-DeepLabV3+ model was improved by 1.19% to 93.48% over DeepLabV3+.The multi-attention module enables the model to integrate multi-scale features more effectively, which not only improves segmentation accuracy but also directly contributes to more reliable and efficient monitoring of tailings ponds. The proposed approach achieves top performance on two benchmark datasets, NX-TPSet and TPSet, demonstrating its effectiveness as a practical and superior method for real-world tailing pond identification.

An analytical model for the in-plane elastic shear modulus of plain-woven fabric composites and its application and experimental validation

In this paper, a new analytical model is proposed that employs experimental data as input to predict the in-plane shear modulus of plain-woven fabric composites. This model boasts high efficiency and rapid computation. Initially, the composite yarn is considered as an idealized curved beam, with its undulation path described by an arc curve and its cross-section simplified into a lenticular shape. Subsequently, a force analysis of the yarn segment is conducted, and the results of the analytical model are derived using the minimum potential energy principle and Castigliano’s second theorem. The input parameters for the analytical model are obtained through thermal ablation experiments and microscopic observations, with T300/Cycom970 plain-woven fabric composites selected as the experimental subject. In addition, tensile tests on ±45° laminates are conducted to validate the effectiveness of the analytical model. The results demonstrate a minimal deviation of −4.15% between the experimental outcomes and the predictions of the analytical model, thereby verifying the model’s validity and the accuracy of the experimental parameters. Leveraging the validated model, the study investigates the influence of crimp ratio and fiber volume fraction on the in-plane shear modulus. Furthermore, to enhance the applicability of the model, a simplified improvement scheme that incorporates temperature and humidity considerations is proposed. This model not only effectively addresses the mechanical property analysis challenges of plain-woven fabric composites, but also provides valuable references for parameter estimation, model construction and experimental design of other woven composite materials.

Automatic visual detection of activated sludge microorganisms based on microscopic phase contrast image optimisation and deep learning.

The types and quantities of microorganisms in activated sludge are directly related to the stability and efficiency of sewage treatment systems. This paper proposes a sludge microorganism detection method based on microscopic phase contrast image optimisation and deep learning. Firstly, a dataset containing eight types of microorganisms is constructed, and an augmentation strategy based on single and multisamples processing is designed to address the issues of sample deficiency and uneven distribution. Secondly, a phase contrast image quality optimisation algorithm based on fused variance is proposed, which can effectively improve the standard deviation, entropy, and detection performance. Thirdly, a lightweight YOLOv8n-SimAM model is designed, which introduces a SimAM attention module to suppress the complex background interference and enhance attentions to the target objects. The lightweight of the network is realised using a detection head based on multiscale information fusion convolutional module. In addition, a new loss function IW-IoU is proposed to improve the generalisation ability and overall performance. Comparative and ablative experiments are conducted, demonstrating the great application potential for rapid and accurate detection of microbial targets. Compared to the baseline model, the proposed method improves the detection accuracy by 12.35% and hastens the running speed by 37.9 frames per second while evidently reducing the model size.

Quantitative Study of Swin Transformer and Loss Function Combinations for Face Anti-Spoofing

Face anti-spoofing (FAS) has always been a hidden danger in network security, especially with the widespread application of facial recognition systems. However, some current FAS methods are not effective at detecting different forgery types and are prone to overfitting, which means they cannot effectively process unseen spoof types. Different loss functions significantly impact the classification effect based on the same feature extraction without considering the quality of the feature extraction. Therefore, it is necessary to find a loss function or a combination of different loss functions for spoofing detection tasks. This paper mainly aims to compare the effects of different loss functions or loss function combinations. We selected the Swin Transformer as the backbone of our training model to extract facial features to ensure the accuracy of the ablation experiment. For the application of loss functions, we adopted four classical loss functions: cross-entropy loss (CE loss), semi-hard triplet loss, L1 loss and focal loss. Finally, this work proposed combinations of Swin Transformers and different loss functions (pairs) to test through in-dataset experiments with some common FAS datasets (CelebA-Spoofing, CASIA-MFSD, Replay attack and OULU-NPU). We conclude that using a single loss function cannot produce the best results for the FAS task, and the best accuracy is obtained when applying triplet loss, cross-entropy loss and Smooth L1 loss as a loss combination.