Cosine Similarity Research Articles

Leveraging paired mammogram views with deep learning for comprehensive breast cancer detection

Employing two standard mammography views is crucial for radiologists, providing comprehensive insights for reliable clinical evaluations. This study introduces paired mammogram view based-network(PMVnet), a novel algorithm designed to enhance breast lesion detection by integrating relational information from paired whole mammograms, addressing the limitations of current methods. Utilizing 1,636 private mammograms, PMVnet combines cosine similarity and the squeeze-and-excitation method within a U-shaped architecture to leverage correlated information. Performance comparisons with single view-based models with VGGnet16, Resnet50, and EfficientnetB5 as encoders revealed PMVnet’s superior capability. Using VGGnet16, PMVnet achieved a Dice Similarity Coefficient (DSC) of 0.709 in segmentation and a recall of 0.950 at 0.156 false positives per image (FPPI) in detection tasks, outperforming the single-view model, which had a DSC of 0.579 and a recall of 0.813 at 0.188 FPPI. These findings demonstrate PMVnet’s effectiveness in reducing false positives and avoiding missed true positives, suggesting its potential as a practical tool in computer-aided diagnosis systems. PMVnet can significantly enhance breast lesion detection, aiding radiologists in making more precise evaluations and improving patient outcomes. Future applications of PMVnet may offer substantial benefits in clinical settings, improving patient care through enhanced diagnostic accuracy.

Improving Paragraph Similarity by Sentence Interaction With BERT

ABSTRACTResearch on semantic similarity between relatively short texts, for example, at word‐ and sentence‐level, has progressed significantly in recent years. However, paragraph‐level similarity has not been researched in as much detail owing to the challenges associated with embedding representations, despite its utility in numerous applications. A rudimentary approach to paragraph‐level similarity involves treating each paragraph as an elongated sentence, thereby encoding the entire paragraph into a single vector. However, this results in the loss of long‐distance dependency information, ignoring interactions between sentences belonging to different paragraphs. In this paper, we propose a simple yet efficient method for estimating paragraph similarity. Given two paragraphs, it first obtains a vector for each sentence by leveraging advanced sentence‐embedding techniques. Next, the similarity between each sentence in the first paragraph and the second paragraph is estimated as the maximum cosine similarity value between the sentence and each sentence in the second paragraph. This process is repeated for all sentences in the first paragraph to determine the maximum similarity of each sentence with the second paragraph. Finally, overall paragraph similarity is computed by averaging the maximum cosine similarity values. This method alleviates long‐range dependency by embedding sentences individually. In addition, it accounts for sentence‐level interactions between the two paragraphs. Experiments conducted on two benchmark data sets demonstrate that the proposed method outperforms the baseline approach that encodes entire paragraphs into single vectors.

SEMbeddings: how to evaluate model misfit before data collection using large-language models

IntroductionRecent developments suggest that Large Language Models (LLMs) provide a promising approach for approximating empirical correlation matrices of item responses by utilizing item embeddings and their cosine similarities. In this paper, we introduce a novel tool, which we label SEMbeddings.MethodsThis tool integrates mpnet-personality (a fine-tuned embedding model) with latent measurement models to assess model fit or misfit prior to data collection. To support our statement, we apply SEMbeddings to the 96 items of the VIA-IS-P, which measures 24 different character strengths, using responses from 31,697 participants.ResultsOur analysis shows a significant, though not perfect, correlation (r = 0.67) between the cosine similarities of embeddings and empirical correlations among items. We then demonstrate how to fit confirmatory factor analyses on the cosine similarity matrices produced by mpnet-personality and interpret the outcomes using modification indices. We found that relying on traditional fit indices when using SEMbeddings can be misleading as they often lead to more conservative conclusions compared to empirical results. Nevertheless, they provide valuable suggestions about possible misfit, and we argue that the modification indices obtained from these models could serve as a useful screening tool to make informed decisions about items prior to data collection.DiscussionAs LLMs become increasingly precise and new fine-tuned models are released, these procedures have the potential to deliver more reliable results, potentially transforming the way new questionnaires are developed.

Dynamic Task Weighting Mechanism for a Task-Aware Approach to Mitigating Catastrophic Forgetting

Catastrophic forgetting is still a big issue in sequential learning and in particular for Natural Language Processing (NLP) models that tend to forget knowledge encoded in previous tasks when learning new targets. To do this, we present a Dynamic Task Weighting Mechanism which forms a part of the Adaptive Knowledge Consolidation (AKC) framework. Our method dynamically adjust knowledge retention to task similarity and task specific performance, while contrasted to static regularization approaches such as Elastic Weight Consolidation (EWC) and Synaptic Intelligence (SI). This mechanism is proposed that involves computing task embeddings with pre-trained models BERT and quantifying their similarity from cosine similarity. To complete the above, we compute a similarity score which is merged with normalized task specific performance metrics of accuracy, F1 score to form an importance score. The model trades adaptability in learning in order to retain previously learned knowledge by prioritizing important tasks and minimizing interference from other unrelated tasks. We show that our proposed mechanism substantially mitigates forgetting and results in accuracy improvements on extensive experiments on standard NLP benchmarks such as GLUE, AG News, and SQuAD. Among baseline methods (EWC, SI, and GEM), the model also has the highest average accuracy of 86.7% and the least amount of forgetting of 6.2%.

Differential Mobility Spectrometry Authenticates the Botanical Origins of Five Monofloral Honeys.

Differential mobility spectrometry (DMS) enables the detection and separation of volatile organic compounds, exploiting differences in ion mobility under varying electric fields, which enhance the resolution of complex mixtures. In this study, DMS, upon integration with chemometrics, enabled the differentiation of the botanical origins of Italian honeys. Utilizing a benchtop differential mobility spectrometer connected to pure air, 219 monofloral honeys of five different botanical origins (acacia, chestnut, citrus, eucalyptus, and linden) were simultaneously analyzed in positive and negative ion modes and with the analysis duration of less than 2 min. Repeatability of the resultant volatilomic fingerprints was verified by cosine similarity with the subsequent elimination of the nonrepeatable data. Afterward, the data for both polarities were concatenated, and an exploratory investigation was carried out into the discrimination capabilities of DMS by principal component analysis. The merged data were then submitted to a statistical analysis for classification. The performance of the resultant random forest (RF) classifier was evaluated by repeated cross-validation and resubstitution into the training set, and finally, this classifier was validated on a withheld subset of data. The performances of the classification model for each test were evaluated by calculating the contingency table-derived parameters of accuracy, Matthew's correlation coefficient (MCC), sensitivity, and specificity. The RF classifier produced high-performance values when predicting the botanical origin of the honey in both training (MCC 84.8%, accuracy 88.0%, sensitivity 87.5%, and specificity 96.9%) and validation subsets (MCC 82.6%, accuracy 86.3%, sensitivity 85.7%, and specificity 96.5%). Some limitations, expected to be mitigated by further research, were encountered in the correct classification of acacia honey.

Improved CKF Collaborative Positioning Algorithm Between UAVs Based on UWB/IMU

Abstract Addressing the challenges of position information solution among multiple Unmanned Aerial Vehicles (UAVs) and the issue of low relative navigation accuracy, an improved collaborative localization algorithm based on the Cubature Kalman Filter (CKF) is proposed. It utilizes a tightly coupled approach to construct a relative ranging model equipped with Ultra-Wideband (UWB) and Inertial Measurement Unit (IMU) sensors, enabling the correction of self-navigation information. The cosine similarity function is utilized to determine the reliability of the observations, and the observation noise matrix is adaptively adjusted based on the new interest vector from each iteration, thereby refining the positioning trajectory for collaborative navigation. Simulation experiments results demonstrate that the improved CKF collaborative localization algorithm significantly enhances the relative positioning accuracy between UAVs, effectively reduces the maximum relative positioning error, and exhibits strong anti-interference capabilities. It is more suitable for indoor multi-UAV and multi-sensor collaborative localization.

Improved SPGD algorithm for optical phased array chip phase error correction in Lidar applications

Random phase errors in silicon-photonic-based OPA (optical phased array) chips can significantly affect their output beam quality, making efficient phase error correction an important requirement for Lidar systems utilizing OPA chip technology. We report an improved ASPGD algorithm for performing OPA random phase error correction with high efficiency and accuracy, especially for large channel count designs. We successfully demonstrated the ASPGD algorithm phase error calibration capability using a fabricated 16-channel chip, and we also showed that the ASPGD algorithm has significantly faster convergence and higher cosine similarity values for the corrected output beam profile compared with traditional SPGD and genetic algorithms, especially for higher channel count OPA chip designs. We believe our work can make an important contribution to the field deployment of OPA technology for chip scale long-range Lidar applications.

Transport-Related Synthetic Time Series: Developing and Applying a Quality Assessment Framework

Data scarcity and privacy concerns in various fields, including transportation, have fueled a growing interest in synthetic data generation. Synthetic datasets offer a practical solution to address data limitations, such as the underrepresentation of minority classes, while maintaining privacy when needed. Notably, recent studies have highlighted the potential of combining real and synthetic data to enhance the accuracy of demand predictions for shared transport services, thereby improving service quality and advancing sustainable transportation. This study introduces a systematic methodology for evaluating the quality of synthetic transport-related time series datasets. The framework incorporates multiple performance indicators addressing six aspects of quality: fidelity, distribution matching, diversity, coverage, and novelty. By combining distributional measures like Hellinger distance with time-series-specific metrics such as dynamic time warping and cosine similarity, the methodology ensures a comprehensive assessment. A clustering-based evaluation is also included to analyze the representation of distinct sub-groups within the data. The methodology was applied to two datasets: passenger counts on an intercity bus route and vehicle speeds along an urban road. While the synthetic speed dataset adequately captured the diversity and patterns of the real data, the passenger count dataset failed to represent key cluster-specific variations. These findings demonstrate the proposed methodology’s ability to identify both satisfactory and unsatisfactory synthetic datasets. Moreover, its sequential design enables the detection of gaps in deeper layers of similarity, going beyond basic distributional alignment. This work underscores the value of tailored evaluation frameworks for synthetic time series, advancing their utility in transportation research and practice.

Performance Analysis of Effective Retrieval of Kannada Translations in Code-Mixed Sentences using BERT and MPnet

Translating Kannada-English (Kn-En) code-mixed text is a challenging task due to the limited availability of Kannada language resources and the inherent complexity of the dataset. This study evaluates the effectiveness of the sentence transformer model, utilizing pre-trained multilingual MPNet and Bidirectional Encoder Representations from Transformers (BERT) architectures, in generating sentence embeddings to enhance translation accuracy. It encodes both code-mixed sentences and their corresponding Kannada translations into high-dimensional embeddings. By employing cosine similarity, it maps input sentences to their closest translations, encoding 2000 code-mixed sentences and their translations using both the MPNet and BERT models. The findings indicate that the MPNet model proved to be more effective, achieving a model accuracy of 98%, compared to BERT's 88%. Moreover, MPNet outperformed BERT in terms of Bilingual Evaluation Understudy (BLEU) and Recall-Oriented Understudy for Gisting Evaluation (ROUGE) scores, attaining 85.0 and 80.0, respectively, while BERT scored 65.3 and 58.7. These results highlight the advanced capabilities of MPNet in translating code-mixed languages and its potential applicability to a broader range of multilingual Natural Language Processing (NLP) tasks.

Enhancing internet of things attack detection using principal component analysis and kernel principal component analysis with cosine distance and sigmoid kernel

The widespread adoption of internet of things (IoT) devices has brought about unprecedented levels of connectivity and convenience. However, it has also introduced significant challenges, particularly in the areas of security and privacy. This study addresses the critical issue of intrusion detection within IoT environments, with a specific focus on analyzing the Iot-23 dataset. Our methodology involves employing principal component analysis (PCA) and kernel PCA for dimensionality reduction. Subsequently, we utilize the k-nearest neighbors (KNN) algorithm for classification purposes. To optimize the performance of the KNN algorithm, we experiment with various feature scaling techniques, such as StandardScaler, MinMaxScaler, and RobustScaler, utilizing different distance metrics. In our analysis, we discovered that employing the cosine distance metric in combination with KNN resulted in superior intrusion detection performance when utilizing PCA. Additionally, when utilizing kernel PCA, we evaluated multiple kernel functions and determined that the radial basis function and sigmoid kernel yielded the most favorable results.

SO-TAD: A surveillance-oriented benchmark for traffic accident detection

Automatic traffic accident detection is a crucial component of Vehicle-to-Infrastructure (V2I) systems and is gaining increasing attention from researchers due to its significant impact on road safety. However, existing datasets often suffer from issues such as data contamination, small scale, and lack of open access, which hinder the advancement of research in this field. To overcome these challenges, we introduce the high-quality, large-scale Surveillance-Oriented Traffic Accident Detection (SO-TAD) dataset, designed with long-tail distribution characteristics. The SO-TAD dataset comprises 2,186 samples, including 282 traffic accident instances, each providing spatio-temporal localization information. Furthermore, within the proposed frame prediction-based traffic accident detection framework, we designed a three-channel scene feature fusion generative adversarial network (TCFF-GAN). This network integrates scene introduction features and scene difference features to generate predictive frame features. Subsequently, traffic accidents are detected based on the stability of the cosine similarity sequence between the predicted frames and the actual frames. We validated the effectiveness of the proposed TCFF-GAN not only through public datasets but also by conducting extensive experiments on the newly introduced SO-TAD dataset. This work also establishes a new benchmark for automatic traffic accident detection research. The code and dataset are available at https://github.com/cccxy-299/so-tad.

Research on virtual collection method of layer house temperature for the construction requirements of digital twin system.

At present, in the context of the highly intensive development of livestock and poultry breeding, digital management is becoming increasingly important, and digital twin systems are gradually being applied. To solve the contradiction between data acquisition and sensor network congestion, a virtual acquisition method based on historical data and real-time reference of point data is proposed when constructing a digital twin system. Firstly, computational fluid dynamics (CFD) simulation was used to analyze and determine the temperature distribution and environmental characteristics inside the layer house, and the collection area was preliminarily divided according to the CFD simulation results. Then, combined with gray correlation degree and cosine similarity analysis, it can effectively identify the reference points highly correlated with the temperature of the key unmonitored area. Finally, WOA was used to optimize the BiLSTM hyperparameters and construct a WOA-BiLSTM virtual acquisition model. It is based on the XGBoost algorithm to determine the actual data collection points, predict the current value based on the actual data of the reference point and the historical data of the test point, and complete virtual collection. Through the test in a farm, the average absolute error between the data of 10 virtual collection points and the actual data was within 0.25 °C, which ensured the reliability of the data. It analyzes the data volume requirements for digital twin modeling and theoretically verifies the supporting role of virtual collection in the construction of digital twin systems.

KFL 작문 자동 채점에서 문서 유사도를 활용한 주제 이탈 텍스트 자동 검출 방안 연구

The purpose of this study is to propose a method for detecting off-topic texts in KFL (Korean as a Foreign Language) automatic writing assessments. To achieve this, a training dataset consisting of 150 texts was constructed using the Korean learner corpus provided by the National Institute of the Korean Language. Based on this dataset, cosine similarity, Euclidean similarity, and Manhattan similarity models were developed. Subsequently, two test datasets were created: one consisting of texts on the same topic as the training dataset (30 texts) and the other consisting of texts on different topics (30 texts). The predictive performance of each model was compared. Analysis using confusion matrix indicated that the cosine similarity model performed the best (accuracy: .98, precision: 1.00, recall: .96, F-1 score: .97), followed by the Euclidean similarity model (accuracy: .81, precision: 1.00, recall: .63, F-1 score: .77). The Manhattan similarity model demonstrated the lowest performance (accuracy: .56, precision: 1.00, recall: .13, F-1 score: .23). These results confirm that using document similarity for the automatic detection of off-topic texts can be practically applicable in actual assessments.

Anti-plagiarism tool helping developers to generate authentic code

Source code plagiarism detection has become a critical area of research with the increasing prevalence of code reuse in academic and professional settings. In order to achieve thorough code comparison, this work introduces a novel tool for detecting source code plagiarism that combines lexical similarity, abstract syntax trees (ASTs) and cosine similarity. The system incorporates a dynamic front-end that was created using Streamlit, providing an intuitive user interface with a code editor that can run code. Through the "Check Similarity" feature, which calculates the plagiarism percentage and finds the most similar file, the application offers real-time plagiarism detection. The methods, benefits, and difficulties of various approaches are examined in this study, with a focus on how well they identify structural and syntactic similarities. The suggested system has a great deal of promise for academic and professional environments, offering reliable and efficient plagiarism detection.

P-765. Enhancing Tuberculosis (TB) Surveillance with Artificial Intelligence: Utilizing Latent Semantic Indexing (LSI) to Automatically Detect TB Cases

Abstract Background Reducing TB surveillance incurs significant costs, as effective control and societal impact minimization rely on robust, ongoing surveillance systems that integrate and analyze comprehensive data for prompt action. Our study supports integrating artificial intelligence (LSI) into TB and other notifiable disease monitoring frameworks to improve trackingFigure 1Binary Matrix of 125 Terms: Presence vs. Absence of Root Keywords Retrieved from the Patient ElectronicRecord for the LSI Automatic Retrieval System.Binary Matrix of 125 Terms: Presence vs. Absence of Root Keywords Retrieved from the Patient Electronic Record for the LSI Automatic Retrieval System. Methods In 2024, we developed a novel application within Datamart A.R.G.U.S.—Assistant for Recovery and Guarding of Urgent and Epidemiological Sentinels—a cloud-based platform hosted on AWS and developed by our team. This application integrates with hospital systems (MV) and laboratory systems (MATRIX and GAL) to enhance the surveillance of healthcare-associated infections (HAI) and notifiable diseases. Data from both hospital and outpatient care are encoded into a binary matrix of 125 terms (root keywords) extracted from patient electronic records. This matrix is then processed using Singular Value Decomposition (SVD) within an LSI-based information retrieval system (Fig. 1). In our model, each patient is represented as a point in a 125-dimensional hyperdimensional space. When projected into three-dimensional space, this framework allows for effective clustering and classification of notifiable Diseases (Fig. 2), specifically TB (Fig. 3).Figure 2:3D Visualization of High-Dimensional Data: Comparing Mandatory Notifiable Diseases versus Non-Notifiable Diseases. Both patient types are well separated in the space, enabling automatic clustering and classification.3D Visualization of High-Dimensional Data: Comparing Mandatory Notifiable Diseases versus Non-Notifiable Diseases. Both patient types are well separated in the space, enabling automatic clustering and classification. Results The LSI-based matrix was constructed using data from January 2023 to February 2024, encompassing a sample of 17,895 cases (lines) across 125 presence versus absence terms, with each case classified as either a notifiable or non-notifiable disease. The 125-dimensional matrix was reduced to 28 factors using SVD. The sensitivity and specificity of the Euclidean distance were better than those of cosine similarity when automatically identifying TB cases in 1,970 independent queries (Fig. 4).Figure 3:3D Visualization of High-Dimensional Data: Comparing Tuberculosis versus Other Mandatory Notifiable Diseases versus Non-Notifiable Diseases. The three types of patients are well separated in the space, enabling automatic clustering and classification.3D Visualization of High-Dimensional Data: Comparing Tuberculosis versus Other Mandatory Notifiable Diseases versus Non-Notifiable Diseases. The three types of patients are well separated in the space, enabling automatic clustering and classification. Conclusion Artificial Intelligence (LSI), significantly enhances the surveillance of TB and other notifiable diseases by enabling real-time, automated case identification up to one day prior (D-1). This advancement allows surveillance personnel to complete their tasks more efficiently, thereby preventing outbreaks and reducing the risk of unforeseen fatalities.Figure 4:Performance Comparison of Cosine vs. Euclidean Distance Metrics in Automatically Identifying TB Cases Using LSI: Euclidean Distance Demonstrates Superior Accuracy and Has Been Integrated into Datamart A.R.G.U.S. for Real-Time Identification of TB and Other Notifiable Diseases as of the Previous Workday (D-1).Performance Comparison of Cosine vs. Euclidean Distance Metrics in Automatically Identifying TB Cases Using LSI: Euclidean Distance Demonstrates Superior Accuracy and Has Been Integrated into Datamart A.R.G.U.S. for Real-Time Identification of TB and Other Notifiable Diseases as of the Previous Workday (D-1). Disclosures All Authors: No reported disclosures

Emotion Detection in Speech: A Deep Learning-Driven Approach Leveraging Acoustic Features within Intelligent Computing Systems

As digital image collections grow in size, there is an increasing need for robust image retrieval systems capable of managing large datasets effectively. This work offerings a novel Content-Based Image Retrieval (CBIR) system designed to enhance retrieval accuracy across both general and medical image datasets. The proposed system leverages U-Net for feature extraction, integrated with an improved weight-learning approach to enhance retrieval performance. A Convolutional neural network, U-Net, a network design famous for its picture segmentation ability, is utilized to capture complex, high-level image features. The approach includes an adaptive, query-aware feature weighting mechanism that applies weight re-scaling to parameterized features, assigning optimized weights to top-ranked images. This CBIR system comprises three main components: image pre-processing, U-Net-based feature extraction, and feature re-weighting. During pre-processing, images undergo augmentation and normalization to increase model robustness. The feature re-weighting process evaluates feature importance using cosine similarity, which further improves discriminative power at retrieval. The proposed CBIR system was tested across various image datasets through extensive experiments, with performance measured in terms of recall, precision and F1-score. The results indicate that integrating feature re-weighting with U-Net-based extraction significantly enhances retrieval effectiveness. This work represents a step forward in developing adaptive image retrieval systems that better respond to diverse retrieval scenarios.

Glam Guidance: Empowering Your Fashion Choices with Image Insights

The paper introduces a Fashion Recommendation System that utilizes computer vision and machine learning to analyze fashion items' visual attributes such as color, texture, and style, offering similar or complementary product recommendations to users. It guides readers through the process of building such a system using Python. Leveraging a pre-trained Convolutional Neural Network (CNN), the system extracts feature from fashion images, enabling calculation of similarities between images for recommendation purposes. Employing cosine similarity, it provides personalized recommendations based on user preferences. By integrating advanced image processing techniques, the system offers users a personalized selection of fashion items, enriching their shopping experience and facilitating style exploration. The system's effectiveness is validated through experiments on a diverse dataset of fashion images, demonstrating its capability to deliver accurate and relevant recommendations. If you are interested in developing a Fashion Recommendation System using image features, this article provides a comprehensive guide.

Potential of Accelerometers to Remotely Early Detect Bovine Ephemeral Fever in Cattle Using Pattern Mining1

Abstract Bovine Ephemeral Fever (BEF), caused by an arthropod-borne rhabdovirus, is widespread in tropical and subtropical regions. It affects cattle with symptoms of fever, lameness, inappetence and in some situations can result in mortality. The goal of this study is to determine if accelerometer data can be used to identify the behaviour patterns that occur when cattle become ill from BEF. Eight heifers in a separate experiment were monitored with 3-axis accelerometers sensors. Movement variation (MV) was calculated from accelerometer data (25 Hz) using 1-minute epochs and then averaged hourly. Two different approaches, cosine similarity (CS) and deviation from previous behavioural patterns, were developed to autonomously detect patterns and recognize the onset of sickness in cattle using accelerometer data. Analyses show that one heifer had behavioural changes one day before the manager observed BEF, and another heifer had behavioural changes on the same day the manager observed BEF. The other six heifers did not display any BEF symptoms. To validate the efficacy of our analytical approaches, we employed them on a separate commercial herd of 73 cows where 4 of the 27 monitored cows were observed with BEF symptoms. Predictions were either on the day or even the day prior to the manager's observation and diagnosis. There were likely no false positives in the first or second trials using the deviation algorithm with sum_deviation formula, but there were several false positives with the other algorithms. These case studies demonstrate the potential of accelerometer data to autonomously detect disease onset, in some cases before it was apparent to the human observer. However, more research is needed to minimize false positives that may occur from other similar diseases, abnormal weather events or cyclical changes in behaviour such as oestrus is required.

Measuring Distances in High Dimensional Spaces Why Average Group Vector Comparisons Exhibit Bias, And What to Do about it

Abstract Analysts often seek to compare representations in high-dimensional space, e.g., embedding vectors of the same word across groups. We show that the distance measures calculated in such cases can exhibit considerable statistical bias, that stems from uncertainty in the estimation of the elements of those vectors. This problem applies to Euclidean distance, cosine similarity, and other similar measures. After illustrating the severity of this problem for text-as-data applications, we provide and validate a bias correction for the squared Euclidean distance. This same correction also substantially reduces bias in ordinary Euclidean distance and cosine similarity estimates, but corrections for these measures are not quite unbiased and are (non-intuitively) bimodal when distances are close to zero. The estimators require obtaining the variance of the latent positions. We (will) implement the estimator in free software, and we offer recommendations for related work.

The Fruit Recognition and Evaluation Method Based on Multi-Model Collaboration

Precision agriculture technology based on computer vision is of great significance in fruit recognition and evaluation. In this study, we propose a fruit recognition and evaluation method based on multi-model collaboration. Firstly, the detection model was used to accurately locate and crop the fruit area, and then the cropped image was input into the classification module for detailed classification. Finally, the classification results were optimized by the feature matching network. In the method, the detection model was based on YOLOv8, and the model was improved by introducing a TripletAttention structure and an Attention Mechanism-Based Feature Fusion (AFM) structure. The improved YOLOv8 model improves the P, R, mAP50, and MAP50-95 indicators by 2.4%, 2.1%, 1%, and 1.3%, respectively, compared with the baseline model on only one generalized “fruit” label dataset. The classification model Swin Transformer used in this study has a classification accuracy of 92.6% on a dataset of 27 fruit categories, and the feature matching network based on cosine similarity can calibrate the classification results with low confidence. The experimental results show that the proposed method can be applied to the maturity assessment of apples and tomatoes, as well as to the non-destructive testing of apples.