Pre-trained Model Research Articles

A novel hybrid CNN-GRU model for website-based user interface code generator

As software engineering techniques evolve, conventional methods become less efficient due to prolonged development cycles and repetitive tasks. There is a growing need for automated systems that can incorporate User Interface (UI) designs into code. This study proposes a novel Hybrid CNN-GRU model for automatically generate website-based UI codes from UI design images. The proposed method consists of a visual model and a language-decoder part. The visual model employs pretrained CNN models, including EfficientNetV2B0, Xception, and EfficientNetV2M, to extract features from images, whereas a GRU serves as the language model and decoder. An empirical study was conducted to evaluate the effectiveness of various hybrid CNN-GRU approaches for generating UI codes. The EfficientNetV2B0 GRU model demonstrated the fastest training time, whereas the Xception GRU model was the most robust for various data complexities. The results of this study will advance the field of automated code generation, simplify the front-end development process, and shorten the time and effort necessitated for UI code implementations.

Invisible backdoor attack with attention and steganography

Recently, with the development and widespread application of deep neural networks (DNNs), backdoor attacks have posed new security threats to the training process of DNNs. Backdoor attacks on neural networks undermine the security and trustworthiness of DNNs by implanting hidden, unauthorized triggers, leading to benign behavior on clean samples while exhibiting malicious behavior on samples containing backdoor triggers. Existing backdoor attacks typically employ triggers that are sample-agnostic and identical for each sample, resulting in poisoned images that lack naturalness and are ineffective against existing backdoor defenses. To address these issues, this paper proposes a novel stealthy backdoor attack, where the backdoor trigger is dynamic and specific to each sample. Specifically, we leverage spatial attention on images and pre-trained models to obtain dynamic triggers, which are then injected using an encoder–decoder network. The design of the injection network benefits from recent advances in steganography research. To demonstrate the effectiveness of the proposed steganographic network, we design two backdoor attack modes named ASBA and ATBA, where ASBA utilizes the steganographic network for attack, while ATBA is a backdoor attack without steganography. Subsequently, we conducted attacks on Deep Neural Networks (DNNs) using four standard datasets. Our extensive experiments show that ASBA surpasses ATBA in terms of stealthiness and resilience against current defensive measures. Furthermore, both ASBA and ATBA demonstrate superior attack efficiency.

SVQ-MAE: an efficient speech pre-training framework with constrained computational resources

Self-supervised learning for speech pre-training models has achieved remarkable success in acquiring superior speech contextual representations by learning from unlabeled audio, excelling in numerous downstream speech tasks. However, the pre-training of these models necessitates significant computational resources and training duration, presenting a high barrier to entry into the realm of pre-training learning. In our efforts, by amalgamating the resource-efficient benefits of the generative learning model, Masked Auto Encoder, with the efficacy of the vector quantization method in discriminative learning, we introduce a novel pre-training framework: Speech Vector Quantization Masked Auto Encoder (SVQ-MAE). Distinct from the majority of SSL frameworks, which require simultaneous construction of speech contextual representations and mask reconstruction within an encoder-only module, we have exclusively designed a decoupled decoder for pre-training SVQ-MAE. This allows the additional decoupled decoder to undertake the mask reconstruction task solely, reducing the learning complexity of pretext tasks and enhancing the encoder’s efficiency in extracting speech contextual representations. Owing to this innovation, by using only 4 GPUs, SVQ-NAE can achieve high performance comparable to wav2vec 2.0, which requires 64 GPUs for training. In the Speech Processing Universal Performance Benchmark, SVQ-MAE surpasses wav2vec 2.0 in both keyword spotting and emotion recognition tasks. Furthermore, in cross-lingual ASR for Mandarin, upon fine-tuning on AISHELL-1, SVQ-MAE achieves a Character Error Rate of 4.09%, outperforming all supervised ASR models.

Pre-trained quantum convolutional neural network for COVID-19 disease classification using computed tomography images

Background The COVID-19 pandemic has had a significant influence on economies and healthcare systems around the globe. One of the most important strategies that has proven to be effective in limiting the disease and reducing its rapid spread is early detection and quick isolation of infections. Several diagnostic tools are currently being used for COVID-19 detection using computed tomography (CT) scan and chest X-ray (CXR) images. Methods In this study, a novel deep learning-based model is proposed for rapid detection of COVID-19 using CT-scan images. The model, called pre-trained quantum convolutional neural network (QCNN), seamlessly combines the strength of quantum computing with the feature extraction capabilities of a pre-trained convolutional neural network (CNN), particularly VGG16. By combining the robust feature learning of classical models with the complex data handling of quantum computing, the combination of QCNN and the pre-trained VGG16 model improves the accuracy of feature extraction and classification, which is the significance of the proposed model compared to classical and quantum-based models in previous works. Results The QCNN model was tested on a SARS-CoV-2 CT dataset, initially without any pre-trained models and then with a variety of pre-trained models, such as ResNet50, ResNet18, VGG16, VGG19, and EfficientNetV2L. The results showed the VGG16 model performs the best. The proposed model achieved 96.78% accuracy, 0.9837 precision, 0.9528 recall, 0.9835 specificity, 0.9678 F1-Score and 0.1373 loss. Conclusion Our study presents pre-trained QCNN models as a viable technique for COVID-19 disease detection, showcasing their effectiveness in reaching higher accuracy and specificity. The current paper adds to the continuous efforts to utilize artificial intelligence to aid healthcare professionals in the diagnosis of COVID-19 patients.

Transforming Image Super-Resolution: A ConvFormer-based Efficient Approach.

Recent progress in single-image super-resolution (SISR) has achieved remarkable performance, yet the computational costs of these methods remain a challenge for deployment on resource-constrained devices. In particular, transformer-based methods, which leverage self-attention mechanisms, have led to significant breakthroughs but also introduce substantial computational costs. To tackle this issue, we introduce the Convolutional Transformer layer (ConvFormer) and propose a ConvFormer-based Super-Resolution network (CFSR), offering an effective and efficient solution for lightweight image super-resolution. The proposed method inherits the advantages of both convolution-based and transformer-based approaches. Specifically, CFSR utilizes large kernel convolutions as a feature mixer to replace the self-attention module, efficiently modeling long-range dependencies and extensive receptive fields with minimal computational overhead. Furthermore, we propose an edge-preserving feed-forward network (EFN) designed to achieve local feature aggregation while effectively preserving high-frequency information. Extensive experiments demonstrate that CFSR strikes an optimal balance between computational cost and performance compared to existing lightweight SR methods. When benchmarked against state-of-the-art methods such as ShuffleMixer, the proposed CFSR achieves a gain of 0.39 dB on the Urban100 dataset for the x2 super-resolution task while requiring 26% and 31% fewer parameters and FLOPs, respectively. The code and pre-trained models are available at https://github.com/Aitical/CFSR.

Deep learning based approach for actinidia flower detection and gender assessment

Pollination is critical for crop development, especially those essential for subsistence. This study addresses the pollination challenges faced by Actinidia, a dioecious plant characterized by female and male flowers on separate plants. Despite the high protein content of pollen, the absence of nectar in kiwifruit flowers poses difficulties in attracting pollinators. Consequently, there is a growing interest in using artificial intelligence and robotic solutions to enable pollination even in unfavourable conditions. These robotic solutions must be able to accurately detect flowers and discern their genders for precise pollination operations. Specifically, upon identifying female Actinidia flowers, the robotic system should approach the stigma to release pollen, while male Actinidia flowers should target the anthers to collect pollen. We identified two primary research gaps: (1) the lack of gender-based flower detection methods and (2) the underutilisation of contemporary deep learning models in this domain. To address these gaps, we evaluated the performance of four pretrained models (YOLOv8, YOLOv5, RT-DETR and DETR) in detecting and determining the gender of Actinidia flowers. We outlined a comprehensive methodology and developed a dataset of manually annotated flowers categorized into two classes based on gender. Our evaluation utilised k-fold cross-validation to rigorously test model performance across diverse subsets of the dataset, addressing the limitations of conventional data splitting methods. DETR provided the most balanced overall performance, achieving precision, recall, F1 score and mAP of 89%, 97%, 93% and 94%, respectively, highlighting its robustness in managing complex detection tasks under varying conditions. These findings underscore the potential of deep learning models for effective gender-specific detection of Actinidia flowers, paving the way for advanced robotic pollination systems.

Digital Grading Company: Deep Neural Networks for Predictive Analytics for Alternative Asset Portfolio Management

In this study, we investigate the application of deep neural networks for predictive analytics in the realm of alternative asset portfolio management, focusing on Trading Card Games (TCGs) like Pokemon. Our primary objective is to understand the optimal trading cards for investment and to determine the appropriate timing for divestment and retention of these assets. Drawing on the successful implementation of machine learning predictive analytics in traditional financial markets, we aim to extend these techniques to alternative investments. To achieve this, we employ a two-pronged approach using a comprehensive dataset of historical sales values for Pokemon TCGs spanning five years, comprising over 200 million unlabeled data points. The first approach utilizes a semi-supervised sequence learning method, where a Long Short-Term Memory (LSTM) model is initially trained as a sequence autoencoder. The pre-trained model's weights are then used in a supervised phase to fine-tune the predictions for future trading card values. The second approach involves heuristic pretraining, which leverages market-derived indicators from existing literature to bridge the gap between traditional feature engineering and deep learning methods. Our technology stack is designed to handle end-to-end data collection, analytics generation, and execution of recommended trading strategies. Simulated backtesting results demonstrate that our Digital Grading Company (DGC) portfolio model outperforms traditional indices, achieving a 357% return compared to the 61% return of the S&P 500 Index over the same period. Real-time analysis further indicates that our model consistently identifies profitable trading opportunities in the TCG market using hourly data and proprietary algorithmic trade logic, executing trades 24/7. The DGC Predictive Portfolio, based on ungraded Pokemon TCG trading cards, also shows superior performance when compared to the Nasdaq 100 Index (Invesco QQQ ETF). Our analytics framework not only predicts the direction of price movement but also provides insights into the magnitude of changes, enabling a more dynamic and responsive trading strategy. The findings suggest that the integration of deep learning models with heuristic indicators has the potential to revolutionize portfolio management in alternative investments, paving the way for more accurate and profitable decision-making in niche markets. This research highlights the promising role of machine learning in enhancing alternative asset management strategies, suggesting future applications in other asset classes. We conclude that deep neural networks, when combined with data-driven heuristics, can significantly enhance trading outcomes, making them valuable tools for investors in the evolving landscape of alternative investments.

On Elastic Language Models

Large-scale pretrained language models have achieved compelling performance in a wide range of language understanding and information retrieval tasks. While their large scales ensure capacity, they also hinder deployment. Knowledge distillation offers an opportunity to compress a large language model to a small one, in order to reach a reasonable latency-performance tradeoff. However, for scenarios where the number of requests (e.g., queries submitted to a search engine) is highly variant, the static tradeoff attained by the compressed language model might not always fit. Once a model is assigned with a static tradeoff, it could be inadequate in that the latency is too high when the number of requests is large, or the performance is too low when the number of requests is small. To this end, we propose an elastic language model ( ElasticLM ) that elastically adjusts the tradeoff according to the request stream. The basic idea is to introduce a compute elasticity to the compressed language model, so that the tradeoff could vary on-the-fly along a scalable and controllable compute. Specifically, we impose an elastic structure to equip ElasticLM with compute elasticity and design an elastic optimization method to learn ElasticLM under compute elasticity. To serve ElasticLM , we apply an elastic schedule. Considering the specificity of information retrieval, we adapt ElasticLM to dense retrieval and reranking, and present an ElasticDenser and an ElasticRanker, respectively. Offline evaluation is conducted on a language understanding benchmark GLUE, and several information retrieval tasks including Natural Question, Trivia QA and MS MARCO. The results show that ElasticLM along with ElasticDenser and ElasticRanker can perform correctly and competitively compared with an array of static baselines. Furthermore, an online simulation with concurrency is also carried out. The results demonstrate that ElasticLM can provide elastic tradeoffs with respect to varying request stream.

Reconstructing representations using diffusion models for multimodal sentiment analysis through reading comprehension

The primary challenge in multimodal sentiment analysis (MSA), which utilizes textual, audio, and visual information to analyze speakers' emotions, lies in constructing representation vectors that incorporate both unimodal semantic and multimodal interaction information. While existing research has extensively focused on multimodal fusion strategies, there remains insufficient exploration of the intrinsic potential within concurrently enhancing unimodal and multimodal representations. To address this gap, we introduce two additional steps to traditional three-step MSA: text modality enhancement through the machine reading comprehension (MRC) framework and multimodal representation reconstruction via proposing the diverse diffusion denoising autoencoder (D3AE) module. The MRC queries are integrated to locate sentiment-related prior knowledge, thereby deepening textual semantic understanding from a pretrained language model. Meanwhile, D3AE employs a single-step denoising strategy along with diffusion models across multiple time intervals, enabling efficient reconstruction and enhancement of multimodal representations. Extensive experiments conducted on two benchmark datasets, CMU-MOSI and CMU-MOSEI, validate that our model, MRC-D3AE, achieves state-of-the-art performance. The superiority of our model over existing baselines is primarily attributed to integrating MRC for enhancing text modality and D3AE for reconstructing multimodal representations.

Carafe enables high quality in silico spectral library generation for data-independent acquisition proteomics.

Data-independent acquisition (DIA)-based mass spectrometry is becoming an increasingly popular mass spectrometry acquisition strategy for carrying out quantitative proteomics experiments. Most of the popular DIA search engines make use of in silico generated spectral libraries. However, the generation of high-quality spectral libraries for DIA data analysis remains a challenge, particularly because most such libraries are generated directly from data-dependent acquisition (DDA) data or are from in silico prediction using models trained on DDA data. In this study, we developed Carafe, a tool that generates high-quality experiment-specific in silico spectral libraries by training deep learning models directly on DIA data. We demonstrate the performance of Carafe on a wide range of DIA datasets, where we observe improved fragment ion intensity prediction and peptide detection relative to existing pretrained DDA models.

Improving the industrial defect recognition in radiographic testing by pre-training on medical radiographs

Deep learning methodologies have gained substantial traction for defect recognition in industrial radiographic testing including welds, castings and other fields. Regardless of the deep learning utilized, it has emerged as a standard configuration to use a model pre-trained from ImageNet to accelerate convergence and enhance recognition accuracy. However, there is a significant gap between the domain of natural images and industrial radiographs, raising the question of whether there might be a superior pre-training method than relying on ImageNet pre-training. Fortunately, medical radiographs are more similar to industrial radiographs than natural images because of the same imaging method. In this paper, we initially utilize numerous medical radiographic images from CheXpert dataset to train a pre-trained CNN model. Then, we apply this model to four distinct tasks within two radiographic testing scenarios to validate its advantages and generalization capabilities. Finally, experiments on multiple datasets indicate that our method brings more benefits than ImageNet pre-training or training from scratch, with a F1 score improvement of 3.41 %–13.72 % for defect classification and a mIoU improvement of 1.05 %–6.58 % for defect segmentation. It demonstrates that pre-training from medical radiographs provides a cost-free improvement for all kinds of tasks in industrial defect recognition.

Biomedical relation extraction method based on ensemble learning and attention mechanism

BackgroundRelation extraction (RE) plays a crucial role in biomedical research as it is essential for uncovering complex semantic relationships between entities in textual data. Given the significance of RE in biomedical informatics and the increasing volume of literature, there is an urgent need for advanced computational models capable of accurately and efficiently extracting these relationships on a large scale.ResultsThis paper proposes a novel approach, SARE, combining ensemble learning Stacking and attention mechanisms to enhance the performance of biomedical relation extraction. By leveraging multiple pre-trained models, SARE demonstrates improved adaptability and robustness across diverse domains. The attention mechanisms enable the model to capture and utilize key information in the text more accurately. SARE achieved performance improvements of 4.8, 8.7, and 0.8 percentage points on the PPI, DDI, and ChemProt datasets, respectively, compared to the original BERT variant and the domain-specific PubMedBERT model.ConclusionsSARE offers a promising solution for improving the accuracy and efficiency of relation extraction tasks in biomedical research, facilitating advancements in biomedical informatics. The results suggest that combining ensemble learning with attention mechanisms is effective for extracting complex relationships from biomedical texts. Our code and data are publicly available at: https://github.com/GS233/Biomedical.

Multimodal Seed Data Augmentation for Low-Resource Audio Latin Cuengh Language

Latin Cuengh is a low-resource dialect that is prevalent in select ethnic minority regions in China. This language presents unique challenges for intelligent research and preservation efforts, primarily due to its oral tradition and the limited availability of textual resources. Prior research has sought to bolster intelligent processing capabilities with regard to Latin Cuengh through data augmentation techniques leveraging scarce textual data, with modest success. In this study, we introduce an innovative multimodal seed data augmentation model designed to significantly enhance the intelligent recognition and comprehension of this dialect. After supplementing the pre-trained model with extensive speech data, we fine-tune its performance with a modest corpus of multilingual textual seed data, employing both Latin Cuengh and Chinese texts as bilingual seed data to enrich its multilingual properties. We then refine its parameters through a variety of downstream tasks. The proposed model achieves a commendable performance across both multi-classification and binary classification tasks, with its average accuracy and F1 measure increasing by more than 3%. Moreover, the model’s training efficiency is substantially ameliorated through strategic seed data augmentation. Our research provides insights into the informatization of low-resource languages and contributes to their dissemination and preservation.

Efficient pretraining model based on multi-scale local visual field feature reconstruction for PCB CT image element segmentation

Element segmentation is a key step in nondestructive testing of printed circuit boards (PCBs) based on computed tomography (CT) technology. In recent years, there has been rapid development of self-supervised pretraining technology that can obtain general image features without labeled samples and then uses a small amount of labeled samples to solve downstream tasks, which has good potential in PCB element segmentation. At present, a masked image modeling (MIM) pretraining model has been initially applied in PCB CT image element segmentation. However, due to the small and regular size of PCB elements such as vias, wires, and pads, the global visual field has redundancy for a single-element reconstruction, which may damage the performance of the model. Based on this issue, we propose an efficient pretraining model based on multi-scale local visual field feature reconstruction for PCB CT image element segmentation (EMLR-seg). In this model, the teacher-guided MIM pretraining model is introduced into PCB CT image element segmentation for the first time, to our knowledge, and a multi-scale local visual field extraction (MVE) module is proposed to reduce redundancy by focusing on local visual fields. At the same time, a simple four-Transformer-blocks decoder is used. Experiments show that EMLR-seg can achieve 88.6% mIoU on the PCB CT image dataset we proposed, which exceeds 1.2% by the baseline model, and the training time is reduced by 29.6 h, a reduction of 17.4% under the same experimental condition, which reflects the advantage of EMLR-seg in terms of performance and efficiency.

A Depth Awareness and Learnable Feature Fusion Network for Enhanced Geometric Perception in Semantic Correspondence.

Deep learning is becoming the most widely used technology for multi-sensor data fusion. Semantic correspondence has recently emerged as a foundational task, enabling a range of downstream applications, such as style or appearance transfer, robot manipulation, and pose estimation, through its ability to provide robust correspondence in RGB images with semantic information. However, current representations generated by self-supervised learning and generative models are often limited in their ability to capture and understand the geometric structure of objects, which is significant for matching the correct details in applications of semantic correspondence. Furthermore, efficiently fusing these two types of features presents an interesting challenge. Achieving harmonious integration of these features is crucial for improving the expressive power of models in various tasks. To tackle these issues, our key idea is to integrate depth information from depth estimation or depth sensors into feature maps and leverage learnable weights for feature fusion. First, depth information is used to model pixel-wise depth distributions, assigning relative depth weights to feature maps for perceiving an object's structural information. Then, based on a contrastive learning optimization objective, a series of weights are optimized to leverage feature maps from self-supervised learning and generative models. Depth features are naturally embedded into feature maps, guiding the network to learn geometric structure information about objects and alleviating depth ambiguity issues. Experiments on the SPair-71K and AP-10K datasets show that the proposed method achieves scores of 81.8 and 83.3 on the percentage of correct keypoints (PCK) at the 0.1 level, respectively. Our approach not only demonstrates significant advantages in experimental results but also introduces the depth awareness module and a learnable feature fusion module, which enhances the understanding of object structures through depth information and fully utilizes features from various pre-trained models, offering new possibilities for the application of deep learning in RGB and depth data fusion technologies. We will also continue to focus on accelerating model inference and optimizing model lightweighting, enabling our model to operate at a faster speed.

Core-Periphery Multi-Modality Feature Alignment for Zero-Shot Medical Image Analysis.

Multi-modality learning, exemplified by the language-image pair pre-trained CLIP model, has demonstrated remarkable performance in enhancing zero-shot capabilities and has gained significant attention recently. However, simply applying language-image pre-trained CLIP to medical image analysis encounters substantial domain shifts, resulting in severe performance degradation due to inherent disparities between natural (non-medical) and medical image characteristics. To address this challenge and uphold or even enhance CLIP's zero-shot capability in medical image analysis, we develop a novel approach, Core-Periphery feature alignment for CLIP (CP-CLIP), to model medical images and corresponding clinical text jointly. To achieve this, we design an auxiliary neural network whose structure is organized by the core-periphery (CP) principle. This auxiliary CP network not only aligns medical image and text features into a unified latent space more efficiently but also ensures alignment driven by principles of brain network organization. In this way, our approach effectively mitigates and further enhances CLIP's zero-shot performance in medical image analysis. More importantly, the proposed CP-CLIP exhibits excellent explanatory capability, enabling the automatic identification of critical disease-related regions in clinical analysis. Extensive experiments and evaluation across five public datasets covering different diseases underscore the superiority of our CP-CLIP in zero-shot medical image prediction and critical features detection, showing its promising utility in multimodal feature alignment in current medical applications.

A hybrid self-supervised model predicting life satisfaction in South Korea.

Life satisfaction pertains to an individual's subjective evaluation of their life quality, grounded in their personal criteria. It stands as a crucial cognitive aspect of subjective wellbeing, offering a reliable gauge of a person's comprehensive wellbeing status. In this research, our objective is to develop a hybrid self-supervised model tailored for predicting individuals' life satisfaction in South Korea. We employed the Busan Metropolitan City Social Survey Data in 2021, a comprehensive dataset compiled by the Big Data Statistics Division of Busan Metropolitan City. After preprocessing, our analysis focused on a total of 32,390 individuals with 51 variables. We developed the self-supervised pre-training TabNet model as a key component of this study. In addition, we integrated the proposed model with the Local Interpretable Model-agnostic Explanation (LIME) technique to enhance the ease and intuitiveness of interpreting local model behavior. The performance of our advanced model surpassed conventional tree-based ML models, registering an AUC of 0.7778 for the training set and 0.7757 for the test set. Furthermore, our integrated model simplifies and clarifies the interpretation of local model actions, effectively navigating past the intricate nuances of TabNet's standard explanatory mechanisms. Our proposed model offers a transparent understanding of AI decisions, making it a valuable tool for professionals in the social sciences and psychology, even if they lack expertise in data analytics.

Enhancing pre-trained models for text summarization: a multi-objective genetic algorithm optimization approach

Enhancing pre-trained models for text summarization: a multi-objective genetic algorithm optimization approach

Advancing remote sensing: a unified deep learning approach with pretrained and custom architectures for high-precision classification

Advancement of remote sensing is vital for accurate land cover mapping and ecological surveillance. This research proposes a cutting-edge deep learning framework that integrates a meticulously designed customary network architecture with the effectiveness of pre-trained models, namely GoogleNet, VGG16, and InceptionV3. Our approach captures the complex features of RGB satellite images throughout various land cover categories using diverse EuroSAT datasets. Our system achieves an impressive 99.40% test accuracy by implementing feature extraction through effective models with a customized deep learning network in an intuitive manner. High class-wise accuracies are observed, ranging from 96.00% to 100.00%, whereas F1-scores, precision, and recall all converge at 99.40%. The aforementioned results demonstrate our approach’s perspective to significantly improve analysis while also confirming its superiority. These results not only confirm the performance of our algorithm but also show that it could be used to redesign satellite visual analysis behavior. These results open the window to further improve and accurate remote sensing systems, giving a helpful viewpoint for researchers as well as professionals.

Field-road classification for agricultural vehicles in China based on pre-trained visual model

Field-road classification that automatically identifies the activity (either in-field or on-road) of each point in Global Navigation Satellite System (GNSS) trajectories is a critical process in the behavior analysis of agricultural vehicles. To capture movement patterns specific to agricultural operations, we propose a multi-view field-road classification method, which extracts a physical and a visual feature vector to represent a trajectory point. We propose a task-specific approach using a pre-trained visual model to effectively extract visual features. Firstly, an image is generated based on a point plus its neighboring points to provide the contextual information of the point. Then, an image recognition model, a fine-tuned ResNet model is developed using the pretraining-finetuning paradigm. In such a paradigm, a pre-training process is used to train an image recognition model (ResNet) with natural image datasets (e.g., ImageNet), and a fine-tuning process is applied to update the parameters of the pre-trained model using the trajectory point images, enabling the model to have both general knowledge and task-specific knowledge. Finally, a visual feature is extracted for a point by the fine-tuned model, thereby overcoming the limitations caused by the small-scale generated images. To validate the effectiveness of our multi-view field-road classification, we conducted experiments on four trajectory datasets (Wheat 2021, Paddy, Wheat 2023, and Wheat 2024). The results demonstrated that the proposed method achieves competitive accuracy performance, i.e., 92.56%, 87.91%, 90.31%, and 94.23% on four trajectory datasets, respectively. Extensive experiments demonstrate that our approach can consistently perform better than the existing state-of-the-art method on the four trajectory datasets by 2.99%, 4.42%, 2.88%, and 2.77% in the F1-score, respectively. In addition, we conduct an in-depth analysis to verify the necessity and effectiveness of our method.