Table 1 Qualitative comparison with existing classification models.

¹

Introduced an

enhanced ResNet-

152 model for bird

classification,

attaining a high

accuracy on the

BIRDS 525 dataset

Improves

classification by

addressing the

vanishing gradients

issue caused by

residual layers

Exhibits difficulties

with occluded

images, which

impact performance

when bird parts are

obscured

²

Utilized

EfficientNetB0 with

data augmentation

and transfer

learning for

classification

Attained an

accuracy of 86.7%,

illustrating the

effectiveness of

transfer learning

Challenging to

classify certain

species due to

some subtle

visual differences

³

Used CNNs with

skip connections

plus VGG16 for

classification

Achieved high

accuracy (92%) for

20 species of South

Indian bird species,

highlighting the

effectiveness of

CNN

Performance dropped

significantly for 525

species, showing

the incapacity of

conventional CNNs

to handle extensive

datasets

⁴

Constructed a

lightweight

attention-

mechanism-based

bird categorization

model based on

ShuffleNetV2

Low processing

complexity and high

accuracy (87.02%)

make it appropriate

for mobile devices

Performance for

fine-grained

classification is

limited by lower

accuracy when

compared to deeper

CNNs

⁵

Bird species were

classified using

MobileNetV2 with

an accuracy of

84.5%

Due to its minimal

computing

requirements, it is

effective for mobile

and edge devices

Faces challenges

classifying data at

high resolution

when larger models

outperform smaller

ones

⁶

Enhanced YOLOv5

for fine-grained bird

classification,

achieving 92%

accuracy

Successfully

manages fine-

grained

classification by

altering the

conventional

YOLOv5

architecture

Challenges with

species that exhibit

significant

intraspecies

variation, which

restricts practical

use

⁷

Provided a Hybrid

Granularities

Transformer for

fine-grained

classification

Enhances

challenging

categorization tasks

by extracting both

local and global

traits

Computationally

demanding, which

makes low-power

or real-time

applications

challenging

⁸

YOLOv5 and

CNNs (VGG19,

Inception V3,

EfficientNetB3) are

combined in this

hybrid model to

recognize and

classify birds

Attained high test

accuracy (98% with

EfficientNetB3),

proving hybrid

architectures’

effectiveness

Struggles with

occluded images,

where only partial

bird images are

visible

⁹

Proposed an

ensemble of fine-

tuned CNN models

for bird

classification

Improved accuracy

with the use of

several pre-trained

networks

It overcomes

occluded images

and overfitting

issues in small

datasets

¹⁰

Presents an

ensemble learning

with DCNNs for

multi-modal

image fusion

Integrates image

data and features

with the help of

ensemble learning,

improving the

accuracy of

species

differentiation

There is little

discussion of direct

application to the

taxonomy of birds

¹¹

Utilized deep

learning ensembles

for multimodal

remote sensing

image classification

Highlights the

potential of

ensembles for

improving

classification in

diverse

environments

Does not focus

on addressing

the computational

complexity of the

task and data

imbalance

¹⁹

Proposed a transfer

learning-based

hierarchical

classification

system for Amazon

parrot species

Achieved high

accuracy (mAP

0.944); useful for

visually similar

species with

insufficient data

Scalability is limited

when introducing

new species or

when there aren’t

many physical

differences

²⁰

Developed SFSCF-

Net for FGVC by

fusing cross-

feature fusion with

saliency

suppression

Enhanced feature

focus and robust

semantic

representation for

fine-grained bird

classification

High computational

complexity, which

precludes its use in

real-time or low-

resource settings