Transformers were initially introduced for natural language processing, leveraging the self-attention mechanism. They require minimal inductive biases in their design and can function effectively as set-based architectures. Additionally, transformers excel at capturing long-range dependencies and enabling parallel processing, which allows them to outperform traditional models, such as long short-term memory (LSTM) networks, on sequence-based tasks. In recent years, transformers have been widely adopted in computer vision, driving remarkable advancements in the field. Previous surveys have provided overviews of transformer applications across various computer vision tasks, such as object detection, activity recognition, and image enhancement. In this survey, we focus specifically on image classification. We begin with an introduction to the fundamental concepts of transformers and highlight the first successful Vision Transformer (ViT). Building on the ViT, we review subsequent improvements and optimizations introduced for image classification tasks. We then compare the strengths and limitations of these transformer-based models against classic convolutional neural networks (CNNs) through experiments. Finally, we explore key challenges and potential future directions for image classification transformers.

Magnetic anomaly signals are composed of anomaly signal and the geomagnetic field. Due to the similarity in magnitude between these two types of signals and the difficulty in acquiring magnetic field data, distinguishing between them is challenging, and the available dataset is small. This paper aims to address the classification of weak magnetic signals with limited samples obtained from actual measurements, a novel neural network-based approach for magnetic anomaly classification is proposed. Firstly, the feature selection is performed on the fused magnetic field signal features. The measured magnetic signals are decomposed using the standard orthogonal basis functions (OBF), and the coefficients of the basis functions are utilized as magnetic moment features. The wavelet transform is employed to calculate the coefficients as the time-frequency features of the magnetic field data. Statistical features are extracted based on the characteristics of the magnetic anomaly data. Using the statistical feature mean as a benchmark, selection is conducted considering the characteristics of the feature dataset, resulting in improved classification results. Afterwards, a lightweight magnetic anomaly classification model, MAD_FA, was designed, resulting in an average reduction of 41.67% in training time. Focal Loss was employed as the loss function during training, leading to an improvement of 3.86% in classification accuracy. A Multi-Feature Adaptive Entropy Weighting(MFAEW) method is proposed to extract magnetic signal features, which adaptively determines feature weights and effectively utilizes the mutual information and complementarity between features. This approach accelerates network convergence and improves classification accuracy by 2.53%. Finally, a comparison was made between the MAD_FA model and classical signal classification models, and a series of ablation experiments were conducted to evaluate the model. The suggested technique performs well in the task of classifying weak magnetic signals, with a classification accuracy of 99.96%, an F1 score of 96.38%, and an AUC score of 99.12%. The higher classification accuracy and stronger robustness compared to the traditional methods demonstrate the potential application of the MAC_FA model in weak magnetic signal classification tasks.

This study aims to maximize the utilization of vector information from small-scale magnetic targets and proposes a high-precision magnetic field reconstruction model based on the magnetic dipole model and the particle swarm optimization(PSO) algorithm. Additionally, it leverages the unique capabilities of the Vision Transformer (ViT) model to effectively handle the characteristics of the reconstructed magnetic anomaly sequence data for classification. To begin with, we establish a magnetic field reconstruction model using the PSO and the magnetic dipole model, introduce the chaotic random inertia weight strategy and calculate the secondary magnetic moment to optimize the model, and suggest the CPSO_MD model. The model can reproduce the recorded magnetic field with excellent precision, and at various observation distances, its reconstruction error is decreased by 3.41% and 2.70%, respectively. In addition, Build domain classifiers to address domain offset and short sample data set issues, and fine-tune the ViT model in accordance with the features of successive magnetic field samples to address the overfitting and oscillation of magnetic anomaly classification issues. Last but not least, the proposed model is evaluated using a range of metrics. The accuracy of the model classification verification is 98.66%, while the accuracy of the area classification verification is 98.21%. Also, the model does not oscillate or overfit during the subsequent rounds of verification, which definitely demonstrates the efficacy of the domain classifier and fine tuning in the model.