Deep Learning Models for Biometric Recognition based on Face, Finger vein, Fingerprint, and Iris: A Survey

No need for physical friction

Easy in keeping templet.

Comfortable, statistics less complicated

Rapid identification procedure

Changes depending on time, age, incidental events,

Differences between twins are difficult.

Affected by lighting in the surrounding environment.

May be partially occluded by other objects

Access control

Face ID

Interaction within computer

Criminal determination

Monitoring

Smart cards

Modern, reliable, safe, highly accurate and less cost

Rapid matching

Need small memory space.

Affected by wound, dust, twists.

Need a physical communication

Authentication of the driver

Criminals' determination and forensics

Authentication in both license and visa cards

Access control

Scalable, accurate and highly covered

Samples of small size

Rapid processing and maximum cost

Have unparalleled structure.

Remains stable throughout the life

Difficult to adjust.

High randomness

No physical contact is needed and just user collaboration.

Hidden by some eye parts such as lashes.

Affected by some illness conditions

Criminals' determination, and forensics

Identification

Access control

National security determining in all of seaports, land, and airports

Sanitary without any touch

Highly accurate and hard to spoof.

Unique

Affected by body temperature.

Affected by some diseases.

Tiny size of template

Minimum processing

Driver identification

Door's security login

Bank services

Physical access monitoring and attendance time

Airports, hospitals, schools

Unsupervised feature learning

Low complexity due to count of parameters and sharing of weights.

High performance in recognition and classification of images

Large dataset required.

Long training time

Unable to deal with input variations (i.e., orientation, position, environment)

Can remember and learn from past data, to give better prediction.

The ability to capture long sequences patterns in the data of large size.

Often utilized for natural language processing tasks

computationally expensive

more porn to overfitting and vanishing gradient problems.

hard to optimize due to the large count of layers and parameters.

Better attitude in dealing with long-term dependencies.

Utilized LSTM cell as activation function so it's less susceptible to the vanishing gradient problem.

Very effective at modeling complex sequential data.

More complicated than RNNs

require more training data in order to learn effectively.

Not suited for prediction or classification tasks.

Slow on large datasets training.

Not work effectively for all kinds of data such as nonlinear or noisy ones.

Uses less memory and is faster than LSTM.

Has fewer parameters than LSTM

low learning efficiency, due to the slow convergence rate

too long training time

may suffer from under-fitting problem

Unsupervised and doesn't need labeled data for training.

Convert the high dimension data into low dimension features.

High scalability with the increase of data.

minimize the noise of entered data

high complexity

computationally expensive,

need large training dataset.

causes losses in interpretability, when representing features in a latent space

Unsupervised feature learning

robust in classification (size, position, color, view angle – rotation).

implemented in many kinds of dataset.

resistant to overfitting due to the RBMs' contribution to model regularization.

Can manage missing data

high complexity

computationally expensive

need large training dataset.

Can deal with partially labelled data.

Efficient generation of samples which looks like the original one.

used in generating images and videos.

Hard to be trained due to the need for different data types in a continuous manner.

training cannot be completed when having missing pattern.

have difficulties in dealing with discrete data (e.g., text)