Research on AI Small Sample Image Recognition

Jixin Chen

1a,*

and Songrong Lv

Wuhan Britain-China School, Wuhan, China

NorthernYucai Foreign Language School, Heping Street, Shenyang, China

Keywords: AI Small Sample Image Recognition, Data Enhancement, Meta Learning.

Abstract: In computer vision and image recognition technology, as the input resolution changes, the recognition effect

of convolutional neural network methods also varies. Meta learning allows computers to simulate the human

brain and learn how to learn, which can achieve image classification more efficiently and flexibly. this paper

will first focus on technical progress, application field expansion and challenge and response of AI image

recognition. This article introduces image classification based on meta learning, followed by image

classification based on data augmentation and image grading based on transfer learning, Finally, a summary

and outlook on small sample image recognition technology were presented. With the continuous develop of

deep learning, transfer learning, meta learning and other technology, AI small sample image recognition

models will be more optimized and able to achieve higher recognition accuracy with less data. The technology

will play an important role in more fields and bring more convenience and value to mankind.

1 INTRODUCTION

There is a significant difference between deep

learning and human intelligence, where humans are

good at identifying new classes of objects with very

small numbers of samples, and deep learning can

easily produce overfitting in this case. Therefore, the

small sample problem has become one of the

important research directions in the field of machine

learning (Ge, Liu, Wang, et al. 2022). Small sample

learning is quite important. Not only can it be used to

learn rare cases. For example, for animals or plants

with scarce information. For many reasons, accurate

and comprehensive data (such as personal privacy or

face data) requires only a small amount of prior

information to accurately classify rare species of

animals and plants. It also reduces the cost of data

collection and calculation, because it requires only a

small amount of data to train the model. Eliminate the

high cost of collecting the data, model training at less

cost. Its application to natural language processing.

Medical applications, acoustic signals processing,

image application. Character recognition and object

recognition and other field, In the field of limited

computing power, such as surveillance cameras and

https://orcid.org/0009-0001-2524-041X

https://orcid.org/0000-0004-3143-9654

unmanned dring, the small sample image recognition

task can well solve the problem because there is not

enough computing power to process so much data.

Nowadays, how to retain the powerful knowledge

representation ability of deep learning while quickly

learning useful knowledge from a small amount of

data for image recognition has become a hot topic, at

present, there are many methods such as meta

learning, transfer learning and data enhancement. In

the following paper is going to introduce several

functions of the technology. Meta learning, also

known as “Learning to learn”, aims to train a “meta

model”, that can quickly adapt and learn new tasks.

Unlike traditional machine learning algorithms,

which focus on training models on specific tasks and

data sets, meta learning focuses more on learning a

general learning strategy or optimization method that

can be applied to a number of different tasks.

Optimization based meta learning focuses on the

optimization process itself, rather than directly

optimizing model parameters. It works by training an

optimizer or learning strategy that enables the model

to quickly find the model, so that it can quickly adjust

and optimize its performance in the face of unknown

territory. MAML is one of the representation

474

Chen, J. and Lv, S.

Research on AI Small Sample Image Recognition.

DOI: 10.5220/0013422200004558

Paper published under CC license (CC BY-NC-ND 4.0)

In Proceedings of the 1st International Conference on Modern Logistics and Supply Chain Management (MLSCM 2024), pages 474-479

ISBN: 978-989-758-738-2

algorithms in meta learning based on optimization

(Gao, Han, Zhang, et al., 2023). MAML is designed

to train a model that is able to adapt rapidly on small

amounts of new task data without having to train from

scratch. The single sample data enhancement method

refers to the processing of a single sample to improve

the generalization ability and performance of the

model.

2 META-LEARNING BASED

METHOD FOR FEW-SHOT

IMAGE CLASSIFICATION

Few-shot image classification based on Meta-

Learning which is the design of models that can adapt

and improve their learning strategies, in other words,

"learning how to learn." It focuses on how to design

algorithms to efficiently learn new tasks with small

amounts of data, thus toward significantly improving

the applicability of the model. This type of learning

method is commonly used by humans, and can also

be used to improve the construction of models that

have the capacity to quickly adapt their parameters to

new data distributions when faced with a new task, to

learn a new task while maintaining the memory of an

already completed task, and to purchase the

knowledge learned on the old task to apply it to the

processing of the new task, thereby improving the

efficiency. In this section, meta-learning methods for

few-shot image classification are discussed through

two meta-learning methods, Model-Agnostic Meta-

Learning and visual meta-learning methods

respectively.

2.1 Model-Agnostic Meta-Learning

That is, to find the best set of parameters which can

be quickly adapted to different models, in order to

achieve the effect of fast convergence with only a

small amount of data references, the model needs a

large amount of prior knowledge to correct this

initialized parameter without interruption, so that it

can be adapted to a large number of different models.

2.2 MAML++

The literature proposes an enhancement method for

MAML, which presents five problems of MAML,

namely, instability during training process,

limitations in model generalization performance,

reduction of flexibility of the framework, high

computational cost, and the need for time-consuming

and labor-intensive parameter tuning of the model

before it can handle a new task, in order to ameliorate

the above problems, the literature proposes

MAML++(Edwards, 2018), which is an improved

MAML-based variant, which provides better

flexibility and more stable training, while improving

efficiency and generalization performance. To solve

the instability problem, the literature implements the

method of optimizing multi-part loss to compute the

target set loss of the minimized base network at each

step towards the support set task, proposing that the

minimized loss is the weighted sum of the target set

loss after each support set loss update, and by using it

to compute the weighted sum, the instability of the

MAML due to backpropagation can be mitigated

efficiently; furthermore, the literature addresses the

issue of high overhead and high cost problem, the

literature proposes annealing the derivative order

during the training process, i.e., using the first-order

gradient to compute the first stage of the training

phase and using the second-order gradient for the rest

of the training phase; to address the problem of the

limitation of the generalization performance, the

literature proposes to learn a different learning rate

and direction for each adaptation process of the

underlying network, which reduces the required

memory and arithmetic while providing flexibility

and generalization performance, and may help to

mitigate overfitting. may help mitigate overfitting.

2.3 Context-Aware Meta-Learning

In recent years, the problem that purely visual domain

models hardly have the ability to detect new objects

when reasoning has gained much attention. To

address this problem, a meta-learning algorithm has

been proposed in the literature. That is, new visual

concepts are learned during inference but not fine-

tuned to simulate large-scale language models, and

meta-learning is redefined as the process of modeling

sequences of data points with known labels and data

points with unknown labels by resorting to a pre-

trained feature extractor.

Figure 1: the process of CAML.

Research on AI Small Sample Image Recognition

475

The general framework of the CAML method is

shown in figure1 and consists of three components: a

CLIP image encoder with frozen parameters, a Fixed

Equal Length and Maximum Equal Angle Set

(ELMES)-like encoder (and a non-causal sequence

model).

(ELMES) class encoder (Fikus, 2018) and a non-

causal sequence model. As shown in the figure 1,

CAML encodes and supports the images using a pre-

trained feature extractor and extracts them into a low-

dimensional representation. The categories of the

support set are encoded using the ELMES category

encoder, but since the category of the query is

unknown, the literature proposes a special learnable

unknown marker embedding. Subsequently, CAML

associates each image embedding with the

corresponding query to form an input sequence that is

significant after large-scale training. In order to

accurately assess the performance of this sub-method

for image recognition classification, the literature

compares it to standard meta-learning evaluation

benchmarks for generic object recognition, fine-

grained image classification, inter-domain image

classification, and unnatural image classification. In

most of the evaluation environments, CAML

performs quite competitively (Clark, 2021).

3 FEW-SHOT IMAGE

CLASSIFICATION BASE ON

DATA AUGMENTATION

The main problem with few-shot image classification

tasks is that the lack of sufficient data, which means

expanding the amount of data by manually processing

the existing data or generating data through

computation is a solution; in other words, just make

some changes to the existing data, such as rotating or

cropping, transposing, etc. to change the data, and

then new data can be obtained. Even the simplest

form of image processing can multiply the data by

several times, and the effect is extraordinary. This

way of making a limited amount of data produce a

value equivalent to more data by not substantially

enhancing the data is data enhancement. this paper

will discuss the methods of single-sample data

enhancement versus multi-sample data enhancement.

3.1 Methods Based on Single-Sample

Data Enhancement

That is, the enhancement of the sample around the

sample itself to carry out operations, including

geometric transformation of the image such as

rotation, cropping, zoom, etc., but also in the color

change or insert noise blurring process.

The literature proposes a flexible data

enhancement approach to vary the representation of

an image without changing its semantics through the

diffusion model DA-fusion, taking into account the

following three conditions: applicability to all images,

minimization of the tuning of a specific dataset, and

the balance between reality and the virtual (Jonathan,

Tim, 2022). The literature proposes two approaches

to prevent Internet data leakage, model-centered

leakage prevention and data-centered leakage

prevention, and also, since standard data

augmentation applies to all images, the literature tries

to keep up with the flexibility of the images through

diffusion-based augmentation techniques to enhance

the credibility of the image generation. The pair

diffuses the model by inserting a new embedding in

the generated text encoder to get a new model that

enhances the credibility of the new image concept.

Figure 2: Comparison chart.

The results are shown in Figure 2, the DA-Fussion

model greatly exceeds the baseline demand and

occupies a competitive position in general, possessing

better results. (Trabucco, 2023)

3.2 Methods for Multi-Sample Data

Enhancement

Unlike single-sample data augmentation, multi-

sample data augmentation utilizes multiple related or

different samples to generate new samples through

computation

The literature is based on the support set X = {(x,

y)} K, N, where x represents an image, while y

represents its label, K represents the number of

images contained in each class, and N denotes the

number of classes. The goal is to learn a function Fθ

(x) → y (which maps an image x to the corresponding

label y). Following the idea of being similar to the

pre-training method and restricting to small sample

parameter training, the paper constructed a new

method DISEF (Diversified In-domain Synthesis

MLSCM 2024 - International Conference on Modern Logistics and Supply Chain Management

476

with Efficient Fine-tuning) based on the foundation

of the pre-trained VLM as shown below in Figure 3:

Figure: 3 SAP process (Victor, et al. 2023).

The Synthetic Augmentation Pipeline (SAP) is used

for fine-tuning, which involves efficient tuning of the

large model.

1. First, the input image (X) is trained by SAP to

produce more training data (X').

2. Second, the images (X) and data (X') are passed

into Vision Encoder to generate visual features (fv);

meanwhile, the class labels are combined with

predefined templates through Text Encoder (T) to

generate text features (ft).

3. Then, logits = sim ( fv, ft ) (sim is a cosine-like

function) and cross-entropy loss Lce are computed.

4. Finally, the initial model is changed by adding

LoRA layers to the query function and embedding the

value (V) of the self-attention (SA) layers in the text

and vision encoder. (Wang, 2023)

The following figure illustrates the SAP process:

1. add labels to the image set (X) using the Image

Description Model (ICM) while projecting them into

the Stable Diffusion latent space.

2. Introduce noise into the latent vector and

randomize the playback using each type of label to

get a synthetic image (X')

3. filter the synthetic images (X') using CLIP to

retain the synthetic images that meet their category

expectations.

4 FEW-SHOT IMAGE

CLASSIFICATION BASED ON

MIGRATION LEARNING

Although machine learning is being applied in an

increasing number of domains, existing supervised

learning requires large amounts of labeled data,

which is both time-consuming and costly, and

migration learning is therefore receiving increasing

attention in order to leverage previously labeled data

to ensure model accuracy for new tasks, in other

words, it is the conversion of knowledge gained from

learning in a domain to learning from existing

datasets in order to improve and enhance the learning

efficiency of the model. Two important concepts are

domain and task; a domain is a specific area, e.g., a

mobile game and a computer game are two different

domains, and a task is something to be done, e.g., a

psychological test and a physical fitness test are two

different tasks. The key points of transfer learning are

to study the shared knowledge that can be transferred

between different domains, to study how to write

specific algorithms to provide the shared knowledge

and enable the transfer, and to study in which cases

transfer is more appropriate and whether the transfer

method is suitable for a particular use case.

4.1 A Transfer Learning Approach

Based on Feature Mapping

Which is focusing on how to find out the common

features between the source and target domains and

mapping them from the original feature space to the

new feature space. The potential shared feature space

found will go to build a bridge between the source and

shared domains. If two domains are related and they

share some data (e.g.the same parameters or common

variables) some of which cause different distributions

but others do not, parameters can be found that do not

affect the distributions to make the distributions of the

source and target domains similar. In this way, the

source domain data in that space is the same as the

target domain data, which leads to better utilization of

the existing labeled data for training in the new space.

In this paper, the paper will develop a discussion on

the mapping based migration learning small sample

image recognition method combined with meta-

learning.

4.2 Sink-Horn Algorithm

The framework of the proposed scheme in the

literature mainly follows the improvements made in

the preprocessing of the latent space output of the

backbone model β, so that the varying eigenvectors

are differently normalized and the results will be

more accurate due to the algorithm operated by the

Sink horn mapping algorithm.

As in figure 4, the literature uses CNN to convert

the image into potential space and transforms the

algorithm to preprocess the vectors and subsequently

processes one of the test data and maps it using the

SINKHOEN algorithm to compare it with the central

domains, which are projected into the potential space

Research on AI Small Sample Image Recognition

477

in the same way as the data that is not yet labeled, and

the closest class will be assigned to be used as a

prediction at the test samples, an algorithm that helps

to conversion of distributions for individual classes to

Gaussian classes.

Figure 4: Mapping Process (Tomáš, Daniel, Pavel, 2021).

The performance of the method was tested

through small sample datasets CIFAR-FS and CUB,

and the method proposed in the literature was

extremely accurate under one to five tests,

outperforming other forms of feature preprocessing,

and achieved the second place in the METADI

competition for the year 2020.

5 CONCLUSION

The transferable knowledge in small sample learning

is highly dependent on the distribution of training

data. If the distribution of training data is significantly

different from that of the new class, the model may

gain biased transmission ability. Therefore, studying

the impact of data bias on model performance is of

great significance. In addition, after summarizing and

analyzing different methods, it was found that better

transferable knowledge can be learned from training

data related to the content. Case intensive related data

and diverse categories can further enrich the diversity

and relevance of the data. With the continuous

development of artificial intelligence deep learning,

small sample recognition will be significantly

improved. Deep learning models, especially

convolutional neural networks and converters, will be

able to better learn and extract effective feature

representations from limited data. The key to small

sample recognition lies in the model's generalization

ability, that is, whether the model can accurately

recognize unseen images when only a small number

of samples are seen. Future research will focus on

developing models with stronger generalization

ability. By introducing techniques such as meta

learning and data augmentation, the model's ability to

adapt to new samples will be improved. The core

challenge of small sample learning lies in the scarcity

of data. To address this challenge.

Researchers need to develop more effective

mechanisms for data collection, tagging, and sharing.

Meanwhile, transfer learning and other techniques

can also be used to transfer models trained on large

datasets to small sample tasks.

In the future, artificial intelligence small sample

recognition technology will play an important role in

many fields, promoting the intelligent transformation

of related industries. With the continuous expansion

of application scenarios, the paper believe that small

sample image recognition technology will achieve

even more brilliant achievements. At the same time,

researchers also need to pay attention to issues such

as data and privacy protection to ensure the healthy

development of technology. In summary, artificial

intelligence small sample image recognition

technology has broad development prospects and

enormous application potential. Future research will

be dedicated to solving technical challenges and

application scenarios, and better serving human

society.

AUTHORS CONTRIBUTION

All the authors contributed equally and their names

were listed in alphabetical order.

REFERENCES

Alec, R., Jong, W. K., Chris, H., Aditya, R., Gabriel, G.,

Sandhini, A., Girish, S., Amanda, A., Pamela, M., Jack,

C., Gretchen, K., Ilya, S., 2021. Learning Transferable

Visual Models From Natural Language Supervision.

arXiv:2103.00020.

Antreas, A., Harrison, E., Amos, S.,2018, How to train your

MAML. arXiv:1810.09502.

Arnav, C., Zhuang, L., Deepak, G. et al. 2023. One-for-all:

Generalized lora for parameter. ArXiv:2306.07967.

Brandon, T., Kyle, D., Max, G., Ruslan, S., 2023. Effective

Data Augmentation with Diffusion Models.

arXiv:2302.07944.

Gao, P., Han, J. M., Zhang, R. R. et al., 2023. Llama-

adapter v2: Parameter-efficient visual instruction model.

ArXiv:2304.15010.

Ge, Y. Z., Liu, H., Wang, Y. et al. 2022. Survey on Deep

Learning Image Recognition in Dilemma of Small

Samples. Software.

MLSCM 2024 - International Conference on Modern Logistics and Supply Chain Management

478

Jonathan, H., Tim, S., 2022. Classifier-free diffusion

guidance. ArXiv:2207.12598.

Matthew, F., John, J., Emily, J.K., Dustin, G.M., 2018.

Equiangular tight frames that contain regular simplices.

Linear Algebra and Its Applications, 98-138.

Tomáš, C., Daniel, V., Pavel, K., 2021, Transfer learning

based few-shot classification using optimal transport

mapping from preprocessed latent space of backbone

neural network. arXiv:2102.05176.

Victor, G. T. C., Nicola, D'A., Yiming, W., Nicu, S., Elisa,

R., 2023, Diversified in-domain synthesis with efficient

fine-tuning for few-shot classification.

arXiv:2312.03046.

Research on AI Small Sample Image Recognition

479