Satriawan, Iqbal Bagus (2026) Analysis of SimAM-EMA Hybrid Attention Module on ResNet-50 for 64×64 Pixel Resolution Image Classification. Undergraduate thesis, UPN Veteran Jawa Timur.
This is the latest version of this item.
|
Text (Cover)
22081010118.-cover-skripsi.pdf Download (1MB) | Preview |
|
|
Text (Bab 1)
22081010118.-bab1-skripsi.pdf Download (264kB) | Preview |
|
![[img]](https://repository.upnjatim.ac.id/style/images/fileicons/text.png) |
Text (Bab 2)
22081010118.-bab2-skripsi.pdf Restricted to Repository staff only until 2 July 2029. Download (1MB) |
|
|
Text (Bab 3)
22081010118.-bab3-skripsi.pdf Restricted to Repository staff only until 2 July 2029. Download (609kB) |
|
|
Text (Bab 4)
22081010118.-bab4.pdf Restricted to Repository staff only until 2 July 2029. Download (2MB) |
|
|
Text (Bab 5)
22081010118.-bab5-skripsi.pdf Download (24kB) | Preview |
|
|
Text (Daftar pustaka)
22081010118.-daftarpustaka-skripsi.pdf Download (245kB) | Preview |
|
|
Text (Lampiran)
22081010118.-lampiran-skripsi.pdf Restricted to Repository staff only until 2 July 2029. Download (385kB) |
Abstract
Attention mechanisms in Convolutional Neural Networks (CNNs) have been shown to improve a model’s ability to emphasize important image features. SimAM (Simple, Parameter-Free Attention Module) is an efficient attention mechanism because it computes attention weights without introducing additional parameters; however, it has limited capability in capturing cross-scale spatial information. In contrast, Efficient Multi-Scale Attention (EMA) processes features at multiple scales and models both channel-wise and spatial relationships more comprehensively. These complementary characteristics suggest that SimAM and EMA can be effectively combined. This study investigates a hybrid SimAM–EMA module implemented sequentially within the bottleneck blocks of ResNet-50. Training stability is improved through differential learning rates and freezing the EMA module during the first three epochs to address initialization differences between the pretrained backbone and the newly added EMA layers. Experiments are conducted on the Tiny ImageNet dataset, which contains 200 classes, using Top-1 accuracy, Top-5 accuracy, and computational complexity as evaluation metrics. The results show that the Hybrid SimAM–EMA model achieves 77.84% Top-1 accuracy and 93.44% Top-5 accuracy, outperforming the Baseline ResNet-50 (76.61%), SimAM-Only (76.23%), and EMA-Only (74.18%) models. In terms of computational efficiency, the parameter count increases by only 0.84% (from 23.92M to 24.12M), while GFLOPs increase by 23.2%. Inference latency rises from 7.97 ms to 20.39 ms per image, mainly due to the computational complexity of EMA rather than parameter growth. Despite this overhead, the Hybrid model provides a Top-1 accuracy gain of 1.23% over the baseline and 3.66% over EMA-Only. Ablation analysis confirms that the performance improvement results from the combination of SimAM’s feature reweighting capability and EMA’s enhanced channel representation learning.
| Item Type: | Thesis (Undergraduate) | ||||||||||||
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| Contributors: |
|
||||||||||||
| Subjects: | Q Science > QA Mathematics > QA75 Electronic computers. Computer science | ||||||||||||
| Divisions: | Faculty of Computer Science > Departemen of Informatics | ||||||||||||
| Depositing User: | Iqbal Bagus satriawan | ||||||||||||
| Date Deposited: | 02 Jul 2026 06:28 | ||||||||||||
| Last Modified: | 02 Jul 2026 06:28 | ||||||||||||
| URI: | https://repository.upnjatim.ac.id/id/eprint/54409 |
Available Versions of this Item
-
Analysis of SimAM-EMA Hybrid Attention Module on ResNet-50 for 64×64 Pixel Resolution Image Classification. (deposited UNSPECIFIED)
- Analysis of SimAM-EMA Hybrid Attention Module on ResNet-50 for 64×64 Pixel Resolution Image Classification. (deposited 02 Jul 2026 06:28) [Currently Displayed]
Actions (login required)
 |
View Item |