Deep Dictionary Learning for Image Recognition with Limited Data

in 2025

K. V. Sai Phani, N. Sumanjali, M. Siva Chinmai, V. Swapna Sree, S. Sowmya and S. Faizunnisa

Department of Computer Science and Engineering, Santhiram Engineering College, NH-40, Nandyal-518501, Kurnool,

Andhra Pradesh, India

Keywords: Deep Dictionary Learning, Image Recognition, Sparse Coding, Limited Data, Transfer Learning,

Semi-Supervised Learning, Data Augmentation, Overfitting, Optimization, Medical Imaging.

Abstract: In case of working with few images available, deep dictionary learning (DDL) is an effective image

recognition method. Hence, DDL combines sparse coding and deep learning to extract meaningful

hierarchical features from images and encourage sparsity in order to improve the computational efficiency.

Traditional deep learning models tend to have more demands on the size of dataset for the effective training

while they do not generalize well when provided with less labeled data, on which DDL can do a good job

utilizing the pre-learned dictionaries and unsupervised learning techniques. It helps in generalizing and

reducing the risk of over fitting. Moreover, DDL is compatible with transfer learning, semi supervised

learning, and data augmentation in order to improve performance with limited data. Moreover, progress in

optimization algorithms and regularization have made DDL models more efficient and stable in the recent

times. When labeled data is hard to obtain, the applications of the technique are vast in fields such as

medical imaging, security surveillance, and autonomous systems. In this paper, we have investigated the

promise of DDL for image recognition with limited data, highlighted benefits of DDL and discussed

difficulties holding back the DDL models training and deployment.

1 INTRODUCTION

In recent years, the deep learning techniques have

played a pivotal role in changing face of image

recognition and tremendously powerful solution has

been offered by them in many applications from

medical imaging to autonomous vehicles. As one of

the major downsides to deploying such models

effectively is the requirement of large amounts of

labeled data, however, often such data is hard to

obtain. In many real world scenarios, especially in

the specialized domains, labeled data is either scarce

or expensive to be collected, and we cannot afford a

deep neural networks that requires huge amount of

the dataset to achieve a good generalization.

Limitation of this approach leads to the exploration

of alternative approaches that can decrease

dependence on large labeled datasets but achieving a

high accuracy.

This problem can be solved using deep

dictionary learning (DDL), which has recently

gained prominence as its solution. DDL encapsulates

and integrates these principles to learn compact,

sparse representations of data in the form of both

efficient and interpretable representations. Usually,

dictionary learning deals with learning a set of basic

functions by which data can be sparsely represented

while deep learning does it in an automatic way by

extracting hierarchical features from raw data. The

synergy between these two methodologies makes

greatest advantage of DDL even when it is fed only

with limited data.

Deep dictionary learning is one of such methods

that learn sparse, compact representations of input

images. By being sparse, this sparsity reduces the

computational burden of the traditional deep learning

models as well as let the model pay more attention to

the most important features while conducting

generalization. When training in a setting where we

have little data we often get into an over fitting

regime, and sparse representations can help such that

the model does not simply memorize spurious

patterns.

The other important strength of DDL is that it is

454

Phani, K. V. S., Sumanjali, N., Chinmai, M. S., Sree, V. S., Sowmya, S. and Faizunnisa, S.

Deep Dictionary Learning for Image Recognition with Limited Data in 2025.

DOI: 10.5220/0013899900004919

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

In Proceedings of the 1st International Conference on Research and Development in Information, Communication, and Computing Technologies (ICRDICCT‘25 2025) - Volume 3, pages

454-461

ISBN: 978-989-758-777-1

compatible with unsupervised learning techniques.

In case there are very few labeled examples we can

use unsupervised learning to pre train the model so

that it learns some useful features from the unlabeled

data. Finally, this pretraining approach can be fine-

tuned using smaller amount of labeled data,

providing a data efficient solution.

Another technique complementary to deep

dictionary learning in low data setting is called

transfer learning. To transfer knowledge from large

scale datasets to small, domain specific datasets,

one can leverage pre trained dictionaries, and neural

network model. This way of transferring knowledge

helps DDL models to adapt rapidly for new task

solely without the necessity of intensive retraining

and reducing the necessary time to deploy. In such

medical imaging domains where labelled data is

difficult to obtain, but there exist publicly available

large datasets whose features can be reused, transfer

learning has proven to work quite well.

Further extending the ability of DDL models to

learn from little data are data augmentation

techniques such as image rotation, flipping, and

scaling. This allows artificially increasing data set

by helping to increase diversity of training set, and

therefore, help the model to generalize to the data

that it has not seen previously. When dealing with

tasks such as image identification, where the

variations in orientation, size, and light intensity

conditions are commonplace, but in any task whose

data is unbalanced or limited, augmentation is

especially useful.

However, there are drawbacks of using deep

dictionary learning for image recognition with

limited data. The main challenge is that it is hard to

achieve a proper balance between the sparsity of the

learned dictionary and the expressed ability to fit

complex patterns. If the dictionary is not too sparse

the model will fail to capture some of the through

needed to be learned, and on the other hand if the

dictionary is not too sparse the benefits of hoping it

are lost.

Finally, deep dictionary learning shows great

potentials for a system to recognize images in a

limited number of labeled data. DDL applies the

strengths of sparse representations, deep learning,

unsupervised learning, transfer learning, and data

augmentation in a fashion that is, to a lesser or

greater degree, data efficient, and useful for

problems in multiple domains, such as medical

imaging, security surveillance, or autonomous

systems. With continual development of this field,

optimization of techniques and architecture of

models will undoubtedly contribute to the growing

performance and applicability of DDL in low data

environments.

2 LITERATURE SURVEY

Zheng, H., Yong, H., & Zhang, L. (2021)

investigated the application of deep convolutional

dictionary learning (DCDL) for image denoising.

They introduced a novel framework that integrates

convolutional neural networks (CNNs) with

dictionary learning techniques, aiming to exploit the

advantages of sparse representations of image

patches. This allows the model to ignore the noise

and keep the most important features of the image.

Indeed, their approach is able to outperform

traditional denoising methods in performance while

also being relevant in disciplines where good image

quality is paramount such as medical imaging or

surveillance. This results in better DCDL

performance due to its ability to retain important

image features despite the presence of noise, which

inevitably leads to improved performance in

downstream tasks like object detection and

recognition.

To address this problem, Zhou et al (2021)

introduce a deep sematic dictionary learning

(DSDL) framework for multi-label image

classification. The proposed model is improved by

introducing the semantic information in the process

of the dictionary learning, which boosted the

accuracy of classifying the images that include more

than one object. Any pixel-wise, object-centric loss

to do localization is replaced with semantic

dictionaries that learn contextual and relational

properties between objects in the image, above

pixel-level features. This capacity to take into

account contextual relationships makes the method

well-suited for applications in domains such as

medical image analysis where images may present

multiple disparate structures (e.g., tumors and

organs) that must be classified simultaneously.

Gao, F., Deng, X., Xu, M., Xu, J. and Dragotti,

P.L. (2022) In: Gao, F., Deng, X., Xu, M., Xu, J. and

Dragotti, P.L. Their approach is a multi-modal

convolutional dictionary learning framework which

integrates different data sources (images, text, and

audio) into the learning process. This allows the

model to adapt to and learn from various kinds of

data, allowing for strong features that may

generalize across different types of data. Overall,

their work suggests that leveraging multiple

modalities can augment the performance of image

recognition models and help address challenges in

Deep Dictionary Learning for Image Recognition with Limited Data in 2025

455

scenarios with rich sensory inputs. This approach is

particularly applicable to the analysis of multimedia

data especially in the context of cross-modal

retrieval systems, where interpretation of

multimodal content contributes significantly to

achieving precise recognition.

Gu, X., Shen, Z., Xue, J., Fan, Y., & Ni, T.

(2021) employed convolutional dictionary learning

with local constraints to brain tumor MR image

classification. Their methodology utilizes the sparse

solution property of dictionary learning in order to

effectively represent major cancer areas within MRI

corpses. By localizing constraints into the model

means focusing only on relevant anatomic subseted

areas, leading to more precise method of detecting

tumors. This approach uses multi-edge graph

segmentation to find better tumor areas, allowing for

higher precision compared to normal image

classification methods and proving useful in

medical imaging, where early detection may

improve patient cure rates.

Yan, R., Liu, Y., Liu, Y., Wang, L., Zhao, R., Bai,

Y., & Gui, Z. (2023) proposed a convolutional

dictionary learning-based approach for denoising

low-dose CT images. Noise in low-dose CT scans

results in images of poorer quality and can lead to

difficulties in achieving accurate diagnoses. This is

the problem that the method used by the authors

addresses; learned with dictionary learning from the

noisy images, this sparse representation allows us to

remove noise while keeping the essential anatomical

pieces. It has shown better results to improve the

quality of low dose CT images compared to other

methods, which is crucial to clinical routines where

reducing the radiation dose is significant. Which is

significantly improving the accuracy of CT scans, a

widely used medical imaging technique.

Khodayar, M., Khodayar, M. E., & Jalali, S. M.J.

(2021) used deep learning for pattern recognition in

photovoltaic energy generation. The dataset was fed

to deep learning models for pattern identification

and prediction of energy generation from

photovoltaic systems. It allows detecting

performance issues and improving reliability, thus

maximizing energy generation. Their work is

important for the renewable energy industry; as

different energy sources are used, it is very

important to manage it properly and maximize

energy generation. To maximize resource

allocation, minimize operation costs, and optimize

energy production, accurate predictive models are

required.

Liu et al. (2021) proposed an autosomal VAE-

based diagnostic one that trains sparse dictionary

learning-based adversarial for wind turbine fault

identification. They introduce a model that merges

the sparsity-inducing characteristics of dictionary

learning with the generative capabilities of VAEs,

enabling the model to discover the sparse

representations of fault signatures in wind turbine

sensor data. The system leverages data analysis to

identify and delineate any potential issues with a

system before they become adverse events, allowing

for proactive maintenance to be performed based,

potentially preventing catastrophic engineering

failures. It encourages early detection of faults,

which helps prevent downtime and avoids repair

costs and increases overall system reliability,

especially in the context of predictive maintenance

in wind energy.

Jiang, Y., & Yin, S. (2023) proposed a new

framework for recognizing heterogeneous-view

occluded facial expression data and named it based

on cycle-consistent adversarial networks

(CycleGAN) and K-SVD dictionary learning. Using

CycleGAN for Data Augmentation: Facial

expressions can often be occluded or incomplete,

leading to inconsistencies in the expression data. K-

SVD dictionary learning is used to ensure that

model is able to learn robust representations in the

presence of occlusions. This type of architecture

could have wide applications in facial recognition

and human-computer interaction, where accurately

identifying facial expressions under hard conditions

is important for effective communication.

Kong, Y., Wang, T., Chu, F., Feng, Z., &

Selesnick, I. (2021). Discriminative dictionary

learning-based sparse classification framework for

machinery fault diagnosis. Because of the content-

rich sensor data that helps isolate faults in machinery

early in the manufacturing process, the model can

extract discriminative features through

discriminative dictionary learning. For instance, this

method is useful to monitor industrial machinery in

real time and is important for detecting faults

immediately to avoid an expensive repair and Pb-

time. It allows for improved overall performance

and reliability of mechanical systems, finding

applications in predictive maintenance and industrial

automation.

Alizadeh, F., Homayoun, H., Batouli, S. A. H.,

Noroozian, M., Sodaie, F., Salari, H. M.,... & Rad,

H.S. (2022) Multi subject dictionary learning for

differential diagnosis of Alzheimer's disease, mild

cognitive impairment and normal subjects using

resting state fMRI data. Only one example involved

the analysis of imaging data, specifically fMRI,

where the authors used dictionary learning to derive

ICRDICCT‘25 2025 - INTERNATIONAL CONFERENCE ON RESEARCH AND DEVELOPMENT IN INFORMATION,

COMMUNICATION, AND COMPUTING TECHNOLOGIES

456

meaningful features and subsequently classify data

to different neurological states. This work has

significant implications for neuroimaging where

diagnosis and accuracy are crucial for treatment and

prevention of diseases. They applied dictionary

learning on multi subject data aiming at enhancing

the diagnosis capacity to classify patients with

Alzheimer’s disease and mild cognitive impairment

versus healthy controls.

3 PROPOSED METHODOLOGY

The proposed methodology leverages deep

dictionary learning (DDL) to address the challenge

of image recognition with limited data. By

combining the strengths of sparse coding and deep

learning architectures, this methodology aims to

improve image recognition accuracy even when the

available labeled data is insufficient. The core

components of the proposed methodology are as

follows (figure 1):

Figure 1: System Architecture.

3.1 Data Preprocessing

Custom to Data Augmentation: To overcome such

trouble of limited data, several data augmentation

techniques are used. These include random rotations,

random scaling, random cropping, flipping and color

jittering. Augmentation brings up the diversity in the

training data making the model generalize better.

Normalization: The image data is normalized to a

certain range to help stabilizing the training and

accelerating the convergence. Generally, the range

can be [0, 1] or [-1, 1].

3.2 Sparse Dictionary Learning

• Traditional methods of dictionary learning

known as K-SVD (K-means Singular Value

Decomposition), where a sparse dictionary

is pre-learned using the data of limited data.

Of course, the model is able to learn some

patterns given very limited data, but that is

with the dictionary as the building blocks of

image features it uses.

• All images can be sparsely represented as a

linear combination of dictionary atoms. The

fact that images are sparsely represented

makes the model be able to focus on the

important features of the image.

• During dictionary learning, L1

Regularization is used to make learned

dictionary sparse. This will prevent

overfitting and ensure that the learned

features are also compact and piecewise

distinct.

3.3 Deep Learning Integration

• Convolutional Neural Networks (CNNs):

A CNN is integrated into the dictionary

learning framework. The image data are

used to automate the learning of

hierarchical feature representations from

the CNN. Then, these learned features are

refined with dictionary learning to learn

low level and high-level patterns.

• Backpropagation is used to train the CNN

end-to-end. The network learns from

training both the optimal weights for the

Deep Dictionary Learning for Image Recognition with Limited Data in 2025

457

CNN layers and sparse representation of

the input images. This gives the model

freedom to make the learned dictionary

adapt to different tasks so as to perform

better in cases of lack of data.

3.4 Transfer Learning

● Since the labeled data is scant, transfer

learning is used and the model is initialized

with weights from a pretrained network for

example a network trained on a large

dataset like ImageNet. This pre-trained

model allows the system to inherit the

learnt knowledge from a broader set of

images.

● First, the pretrained model is fine-tuned

using the limited dataset. The learned

dictionary is updated together with the

CNN parameters during fine tuning stage,

the model is updated to the specific image

recognition task without losing universality

of the learned features of the pretrained

model.

3.5 Multi-Scale Feature Fusion

● Feature Fusion: Multi scale feature fusion

is used to improve the model capability to

detect complex patterns. At different scales

the features are extracted and combined

together to form a rich representation of the

image. In doing so this approach allows us

to capture both fine details as well as

broader contextual information out of the

images.

● Fusion Layer: Multi scale features are

combined by adding a fusion layer after the

convolutional layers. The extra layer helps

to increase the model’s power to predict

more accurately by using information from

different scales.

3.6 Post-Processing and Classification

● Dictionaries are learned via Deep

Dictionary Learning model, where after

learning the sparse representation the

model is used for obtaining the most

relevant features using feature selection

methods (e.g. principal component analysis,

mutual information). In order to decrease

the dimensionality of feature space and

enhance the classification performance, this

reduces the number of features.

● The classifier is a fully connected layer or a

support vector machine (SVM) which

performs the prediction using the sparse

feature representation. The selected features

are trained by the classifier and optimized

to enhance accuracy.

3.7 Evaluation and Fine-Tuning

● Cross validation is used to evaluate

performance of the proposed model. The

dataset is small so this means that the

dataset is split into number of training and

validation subsets since the model does not

overfit to the training data.

● Hyper Parameter Tuning: Some hyper

parameters like learning rate, dictionary size

and some regularization parameter(s) differ

hence, the model is fine-tuned with those.

In order to achieve the best performance on

the validation set, we use a grid search or a

random search method to find the best set

of hyper parameters.

3.8 Model Deployment

● Inference: The model is deployed for

inference. In the case of a given test image,

the sparse representation is computed by

the learned dictionary, the relevant features

are extracted by the CNN and a final

classification decision is made based on the

classifier.

● It can be used in real time applications such

as medical imaging, security surveillance,

or industrial monitoring where the new

images or frames are continuously

processed to get recognized or anomalous.

The proposed methodology is a process of

improving image recognition through a combination

of deep dictionary learning and deep convolutional

neural networks using a small amount of data. It

resolves the problem of overfitting by using the

sparse representation and pretraining techniques and

improves its generalization model. The key

components of the methodology, that is, data

augmentation, transfer learning, multi scale feature

fusion and post processing allow robust performance

even when the size of the dataset is small. Thus, this

hybrid approach shows a promising solution for the

image recognition problems in the domain of

constrained data availability.

ICRDICCT‘25 2025 - INTERNATIONAL CONFERENCE ON RESEARCH AND DEVELOPMENT IN INFORMATION,

COMMUNICATION, AND COMPUTING TECHNOLOGIES

458

4 RESULTS AND DISCUSSION

4.1 Experimental Setup

4.1.1 Dataset

The model is evaluated on CIFAR-10 dataset (10

classes) with 60,000 images. Each class contains

6,000 images. Because it considers the limited data

scenarios, we generated subset of the data set. With

varying numbers of labeled images to simulate real-

world situations where labeled data is scarce.

4.1.2 Implementation Details

● We use ImageNet trained CNN (ResNet-

50) as pretrained model and fine tune it

with CIFAR-10 dataset

● Data Augmentation: Random rotation,

flipping and color jittering are applied to

augment the dataset.

● For training the model, the batch size is

fixed to 32 and a learning rate of 0.001 is

used for training the model in for 50

epochs. The dimension of the dictionary

is fixed to 256 atoms.

4.2 Quantitative Results

The performance of the proposed deep dictionary

learning model is compared with traditional deep

learning models (i.e., standard CNNs, ResNet-50

without dictionary learning, and SVM with

handcrafted features).

4.2.1 Classification Accuracy

After training on subsets of the dataset with limiting

labeled images, a measure of the classification

accuracy was made of the test set. The results show

that deep dictionary learning overcomes the

limitations of small number of labeled images

through the benefits it entails in reducing

classification error.

Observation: The proposed model (DDL +

ResNet-50) consistently outperforms the baseline

CNN and even ResNet-50 fine-tuned without

dictionary learning, especially in limited data

scenarios. This confirms that the integration of

dictionary learning with deep neural networks

enhances feature extraction, leading to higher

accuracy even when only a small fraction of labeled

data is available. Table 1 presents the classification

accuracy for each model with varying numbers of

labeled training images.

Table 1: Classification accuracy for each model with

varying numbers of labeled training images.

Number of

Labeled

Images

CNN

(Baseli

ne)

ResNet-50

(Fine-tune

DDL +

ResNet-50

(Proposed)

100 58.3% 62.5% 72.4%

500 71.5% 74.1% 81.2%

1,000 77.9% 79.8% 85.6%

5,000 83.6% 84.2% 90.3%

Full Dataset

(50,000)

89.7% 90.5% 94.1%

4.2.2 Sparsity of the Learned Dictionary

To assess the efficiency of the learned dictionary, we

examine the sparsity of the dictionary learned by the

model. The dictionary is learned using K-SVD, and

the sparsity is controlled by applying L1

regularization. The degree of sparsity is measured by

the percentage of dictionary atoms with non-zero

coefficients in the learned representation.

Figure 2 shows the sparsity of the dictionary for

different number.

Figure 2: Sparsity of the Learned Dictionary.

● When the number of labeled training

images increases, the dictionary becomes

denser, it is still highly sparse compared to

existing deep learning methods.

● This implies that the smaller dictionary

learnt by the model with fewer labeled

images is able to represent images better

with a smaller set of basic functions.

Deep Dictionary Learning for Image Recognition with Limited Data in 2025

459

4.3 Qualitative Results

We visualize some of the learned sparse

representations to better understand how the model

recognizes key features of images.

Figure 3 presents example patches from the learned

dictionary when trained on a small subset of the

dataset (100 labeled images). The learned dictionary

atoms correspond to fundamental image components

such as edges, textures, and simple shapes, which

the model uses to reconstruct the input images

sparsely.

Figure 3: Visualized Dictionary Atoms from DDL.

Each column represents a learned atom in the

dictionary, showing the fundamental features that

the model has learned from the limited data. This

demonstrates how the sparse representation captures

essential patterns despite the limited labeled data.

4.4 Comparison with Traditional

Methods

To validate effectiveness of proposed approach,

performance of the model is compared and

contrasted with a trained deep neural network as

well as the standard methods of image recognition,

that is, Support Vector Machines (SVM). For

classification, we use histogram of oriented gradient

(HOG) and local binary pattern (LBP) as

handcrafted features in baseline SVM method.

Figure 4 shows the classification accuracy of the

proposed method is compared with SVM and a fully

trained CNN on the CIFAR-10 dataset. Obviously,

the proposed method surpasses SVM and CNN with

much richer data; however, with limited data, it also

consistently outperforms both.

Figure 4: Comparison of Classification Accuracy for

Different Models.

The proposed DDL + ResNet-50 model shows

higher accuracy across all data sizes, particularly in

the limited data regime. This reinforces the

advantage of dictionary learning in feature extraction

for limited data tasks.

4.5 Discussion

The results indicate that deep dictionary learning

provides a robust solution for image recognition in

cases where labeled data is limited. The integration

of dictionary learning with deep neural networks

improves feature extraction, enabling the model to

generalize better from fewer labeled examples. The

sparsity induced by the dictionary learning process is

particularly beneficial in limiting overfitting, making

the model more efficient when working with small

datasets.

● Sparsity: The learned dictionary helps in

maintaining high sparsity, ensuring that the

model captures only the most important

features of the data, which is crucial for

limited data tasks. This sparse representation

aids in efficient training and reduces the risk

of overfitting.

● Effectiveness in Limited Data: The

proposed methodology significantly

outperforms traditional deep learning

approaches in scenarios with limited labeled

data. The model's performance increases as

the amount of labeled data grows, but it

consistently outperforms other methods even

when the dataset is small, demonstrating its

robustness in real-world applications.

● Future Work: Future improvements could

involve further optimization of the dictionary

learning process, possibly by exploring

different regularization techniques or

ICRDICCT‘25 2025 - INTERNATIONAL CONFERENCE ON RESEARCH AND DEVELOPMENT IN INFORMATION,

COMMUNICATION, AND COMPUTING TECHNOLOGIES

460

incorporating unsupervised pretraining

methods to further enhance model

performance in low-data environments.

Moreover, expanding the methodology to

work with larger datasets and more complex

architectures could yield even better results.

5 CONCLUSIONS

In this work, we proposed a deep dictionary learning

(DDL) approach for image recognition in scenarios

with limited labeled data. By combining sparse

representation techniques with deep convolutional

neural networks (CNNs), the model effectively

improves feature extraction and classification

accuracy, especially when the amount of labeled

data is limited. Our experiments demonstrated that

the DDL + ResNet-50 model outperforms both the

baseline CNN and the fine-tuned ResNet-50,

particularly in scenarios with few labeled images.

Additionally, the learned dictionary maintains a high

level of sparsity, ensuring that the model focuses on

the most important features, which enhances its

efficiency and generalization capabilities. The

proposed methodology proves to be highly effective

for applications like medical imaging and

surveillance, where labeled data can be scarce. In

conclusion, the approach provides a robust solution

for image recognition tasks with limited labeled

data, and future work will focus on refining the

learning process and extending the model to more

complex image recognition challenges.

REFERENCES

Alizadeh, F., Homayoun, H., Batouli, S. A. H., Noroozian,

M., Sodaie, F., Salari, H. M., & Rad, H. S. (2022).

Differential Diagnosis among Alzheimer's disease,

Mild Cognitive Impairment, and Normal Subjects

Using Resting-State fMRI Data Extracted from Multi-

Subject Dictionary Learning Atlas: A Deep Learning-

Based Study. Frontiers in Biomedical Technologies.

Gao, F., Deng, X., Xu, M., Xu, J., & Dragotti, P. L.

(2022).Multi-modal convolutional dictionary learning.

IEEE Transactions on Image Processing, 31, 1325-

1339.

Gu, X., Shen, Z., Xue, J., Fan, Y., & Ni, T. (2021).

Brain tumor MR image classification using

convolutional dictionary learning with local

constraint. Frontiers in Neuroscience, 15, 679847.

Yan, R., Liu, Y., Liu, Y., Wang, L., Zhao, R., Bai, Y., &

Gui,Z. (2023). Image denoising for low-dose CT via

convolutional dictionary learning and neural network.

IEEE Transactions on Computational Imaging, 9, 83-

93.

Khodayar, M., Khodayar, M. E., & Jalali, S. M. J. (2021).

Deep learning for pattern recognition of photovoltaic

energy generation. The Electricity Journal, 34(1),

106882.

Liu, X., Teng, W., Wu, S., Wu, X., Liu, Y., & Ma, Z.

(2021).Sparse dictionary learning based adversarial

variational auto-encoders for fault identification of

wind turbines. Measurement, 183, 109810.

Jiang, Y., & Yin, S. (2023). Heterogenous-view occluded

expression data recognition based on cycle-consistent

adversarial network and K-SVD dictionary learning

under intelligent cooperative robot environment.

Computer Science and Information Systems, 20(4),

1869-1883.

Kong, Y., Wang, T., Chu, F., Feng, Z., & Selesnick, I.

(2021). Discriminative dictionary learning-based

sparse classification framework for data-driven

machinery fault diagnosis. IEEE Sensors Journal,

21(6), 8117-8129.

Zheng, H., Yong, H., & Zhang, L. (2021). Deep

convolutional dictionary learning for image denoising.

In Proceedings of the IEEE/CVF conference on

computer vision and pattern recognition (pp. 630-

641).

Zhou, F., Huang, S., & Xing, Y. (2021, May). Deep

semantic dictionary learning for multi-label image

classification. In Proceedings of the AAAI conference

on artificial intelligence (Vol. 35, No. 4, pp. 3572-

3580).

Deep Dictionary Learning for Image Recognition with Limited Data in 2025

461