Conventional image classification methods usually require a large number of training samples for the training model. However, in practical scenarios, the amount of available sample data is often insufficient, which easily leads to overfitting in network construction. Few-shot learning provides an effective solution to this problem and has been a hot research topic. This paper provides an intensive survey on the state-of-the-art techniques in image classification based on few-shot learning. According to the different deep learning mechanisms, the existing algorithms are divided into four categories: transfer learning based, meta-learning based, data augmentation based, and multimodal based methods. Transfer learning based methods transfer useful prior knowledge from the source domain to the target domain. Meta-learning based methods employ past prior knowledge to guide the learning of new tasks. Data augmentation based methods expand the amount of sample data with auxiliary information. Multimodal based methods use the information of the auxiliary modal to facilitate the implementation of image classification tasks. This paper also summarizes the few-shot image datasets available in the literature, and experimental results tested by some representative algorithms are provided to compare their performance and analyze their pros and cons. In addition, the application of existing research outcomes on few-shot image classification in different practical fields are discussed. Finally, a few future research directions are identified.

The combination of features from the convolutional layer and the fully connected layer of a convolutional neural network (CNN) provides an effective way to improve the performance of crime scene investigation (CSI) image classification. However, in existing work, as the weights in feature fusion do not change after the training phase, it may produce inaccurate image features which affect classification results. To solve this problem, this paper proposes an adaptive feature fusion method based on an auto-encoder to improve classification accuracy. The method includes the following steps: Firstly, the CNN model is trained by transfer learning. Next, the features of the convolution layer and the fully connected layer are extracted respectively. These extracted features are then passed into the auto-encoder for further learning with Softmax normalisation to obtain the adaptive weights for performing final classification. Experiments demonstrated that the proposed method achieves higher CSI image classification performance compared with fix weights feature fusion.

Human action recognition has drawn significant attention because of its importance in computer vision-based applications. Action recognition based on skeleton sequences has rapidly advanced in the last decade. Conventional deep learning-based approaches are based on extracting skeleton sequences through convolutional operations. Most of these architectures are implemented by learning spatial and temporal features through multiple streams. These studies have enlightened the action recognition endeavor from various algorithmic angles. However, three common issues are observed: (1) The models are usually complicated; therefore, they have a correspondingly higher computational complexity. (2) For supervised learning models, the reliance on labels during training is always a drawback. (3) Implementing large models is not beneficial to real-time applications. To address the above issues, in this paper, we propose a multi-layer perceptron (MLP)-based self-supervised learning framework with a contrastive learning loss function (ConMLP). ConMLP does not require a massive computational setup; it can effectively reduce the consumption of computational resources. Compared with supervised learning frameworks, ConMLP is friendly to the huge amount of unlabeled training data. In addition, it has low requirements for system configuration and is more conducive to being embedded in real-world applications. Extensive experiments show that ConMLP achieves the top one inference result of 96.9% on the NTU RGB+D dataset. This accuracy is higher than the state-of-the-art self-supervised learning method. Meanwhile, ConMLP is also evaluated in a supervised learning manner, which has achieved comparable performance to the state of the art of recognition accuracy.

Tyre pattern image retrieval (TPIR) is an important tool in the investigation of criminal activities and traffic accidents. Although content-based image retrieval (CBIR) has been developed for decades with abundant results, the study on TPIR which started in the 1990s has not made much progress. The lack of large standard test datasets is a crucial shortcoming which limits the research in this field. Information presented in this paper is a result of the authors’ literature research on recent academic publications and practical field investigation in the public security and transportation sectors. The state-of-the-art technologies in the field of TPIR are surveyed in detail from two aspects of tyre patterns – their low-level spatial features and high-level semantic features. Existing algorithms are examined and their pros and cons are compared and verified through experimental results. This paper also surveys the available tyre pattern datasets used in all available literature. Finally, with the considerations on technology trends in image retrieval and application requirements in TPIR, the future research directions in this field are laid out.

Fine‐grained image classification (FGIC) aims to distinguish visually similar categories by capturing subtle differences, yet the coexistence of large intra‐class variation and small inter‐class differences makes this task highly challenging. This paper provides a systematic review of recent deep learning‐based FGIC methods. According to the type of training data, existing approaches are categorized into four groups: (1) conventional models with large‐scale samples (strongly supervised, weakly supervised, semi‐supervised, and unsupervised); (2) few‐shot learning models for limited‐sample scenarios (e.g. meta‐learning and metric learning); (3) models leveraging external information, including multi‐modal and web‐sourced data; and (4) emerging diffusion‐based models. Representative algorithms in each category are summarized and analysed in terms of their advantages and limitations. The paper also reviews mainstream benchmark datasets and introduces a newly proposed application‐oriented dataset, CIIP‐TPID, to support real‐world tasks. Additionally, practical applications of FGIC in public security, medicine, and commerce are discussed. Finally, future research directions are outlined, including diffusion‐based data augmentation, advanced multi‐modal fusion, transformer architecture optimization, lightweight models for edge deployment, and robustness against noisy labels. This review provides a structured and up‐to‐date reference for researchers and practitioners in the field.