Lightweight inception-UNet with attention mechanisms for semantic segmentation

Twinkle Tiwari; Mukesh Saraswat

doi:10.2478/ijssis-2026-0014

.blurhash-client-img { display: none !important; }

Lightweight inception-UNet with attention mechanisms for semantic segmentation

International Journal on Smart Sensing and Intelligent Systems

Volume 19 (2026): Issue 1 (January 2026)

By: Twinkle Tiwari and Mukesh Saraswat

Open Access

|Apr 2026

Figures & Tables

Architecture of the encoder block of the proposed LWA-ModUNet.

Architecture of the inception block of the proposed LWA-ModUNet [54].

Decoder architecture of the proposed LWA-MoDUnet.

Diagram of additive attention gate [55].

Flowchart of the proposed LWA-MoDUNet depicting the encoding and decoding phases with inception and attention modules.

Sample images of considered datasets (Row 1) along with their corresponding GT (Row 2). GT, ground truth; IDD-Lite, Indian driving dataset-lite; MADS, martial arts, dancing and sports.

Table 1:

Ablation study of LWA-MoDUNet on autorickshaw dataset

Measure	LWA-MoDUNet	W/o inception module	W/o attention
IoU score	0.8768	0.6729	0.8173
Error	0.0663	0.1452	0.0927
Accuracy	93.3691	78.7966	86.4217
F-score	0.9344	0.7913	0.8786

[i] IoU, intersection over union.

Generated output masks by UNet, E-Net, SegNet, UNet with ResNet-18 encoder, UNet with ResNet-34 encoder, and LWA-ModUNet models on exemplary images of autorickshaw dataset. GT, ground truth.

Generated output masks by UNet, E-Net, SegNet, UNet with ResNet-18 encoder, UNet with ResNet-34 encoder, and LWA-ModUNet on exemplary images of IDD Lite dataset. GT, ground truth; IDD-Lite, Indian driving dataset-lite.

Table 2:

Comparison of key performance indicators of LWA-MoDUNet and SOTA models on autorickshaw dataset

Measures	UNet	UNet-ResNet34	UNet-ResNet18	E-Net	SegNet	LWA-MoDUNet
IoU score	0.7665	0.8459	0.8356	0.8480	0.7592	0.8768
Err	0.1326	0.0849	0.0912	0.0857	0.1403	0.0663
Acc	86.7412	91.5061	90.8837	91.4349	85.9677	93.3691
Sp	0.8697	0.9303	0.9241	0.9514	0.8789	0.9429
Ss	0.8651	0.9009	0.8946	0.8829	0.8423	0.9248
F-score	0.8678	0.9165	0.9104	0.9177	0.8631	0.9344
Cc	0.7348	0.8306	0.8182	0.8315	0.7203	0.8676

[i] Acc, accuracy; Cc, correlation coefficient; Err, error; IoU, intersection over union; Sp, specificity; Ss, sensitivity.

Table 3:

Comparison of key performance indicators of LWA-MoDUNet and compared models on IDD-Lite dataset

Measures	UNet	UNet-ResNet34	UNet-ResNet18	E-Net	SegNet	LWA-MoDUNet
IoU score	0.6031	0.6174	0.5981	0.566	0.3076	0.9283
Acc	0.9203	0.9398	0.9356	0.9321	0.8971	0.9616
Err	0.0716	0.0601	0.0643	0.0679	0.1028	0.00435
Sp	0.9500	0.9484	0.9469	0.9395	0.8975	0.9328
Ss	0.8534	0.8617	0.8472	0.8669	0.8896	0.9436
F-score	0.7056	0.7635	0.7485	0.7229	0.4705	0.8909
Cc	0.6794	0.7371	0.7186	0.6979	0.4965	0.8153

[i] Acc, accuracy; Cc, correlation coefficient; Err, error; IDD-Lite, Indian driving dataset-lite; IoU, intersection over union; Sp, specificity; Ss, sensitivity.

Table 4:

Comparison of key performance indicators of LWA-MoDUNet and compared models on PETS dataset

Measures	UNet	UNet-ResNet34	UNet-ResNet18	E-Net	SegNet	LWA-MoDUNet
IoU score	0.7665	0.8459	0.8356	0.8480	0.7592	0.8768
Err	0.1326	0.0849	0.0912	0.0857	0.1403	0.0663
Acc	86.7412	91.5061	90.8837	91.4349	85.9677	93.3691
Sp	0.8697	0.9303	0.9241	0.9514	0.8789	0.9429
Ss	0.8651	0.9009	0.8946	0.8829	0.8423	0.9248
F-score	0.8678	0.9165	0.9104	0.9177	0.8631	0.9344
Cc	0.7348	0.8306	0.8182	0.8315	0.7203	0.8676

[i] Acc, accuracy; Cc, correlation coefficient; Err, error; IoU, intersection over union; Sp, specificity; Ss, sensitivity.

Table 5:

Comparison of key performance indicators of LWA-MoDUNet and compared models on MADS dataset

Measures	UNet	UNet-ResNet34	UNet-ResNet18	E-Net	SegNet	LWA-MoDUNet
IoU score	0.8968	0.9448	0.9084	0.9609	0.9721	0.9768
Err	0.0545	0.0290	0.0496	0.0201	0.0142	0.0118
Acc	94.5520	97.1042	95.0428	97.9919	98.5781	98.8225
Sp	0.9464	0.9905	0.9818	0.9872	0.9916	0.9934
Ss	0.9447	0.9531	0.9229	0.9728	0.9801	0.9832
F-score	0.9456	0.9716	0.9520	0.9801	0.9859	0.9883
Cc	0.8910	0.9428	0.9028	0.9599	0.9716	0.9765

[i] Acc, accuracy; Cc, correlation coefficient; Err, error; IoU, intersection over union; MADS, martial arts, dancing and sports; Sp, specificity; Ss, sensitivity.

Table 6:

Comparison of key performance indicators of LWA-MoDUNet and compared models on CT-Liver dataset

Measures	UNet	UNet-ResNet34	UNet-ResNet18	E-Net	SegNet	LWA-MoDUNet
IoU score	0.8839	0.9445	0.9565	0.9349	0.9495	0.9650
Err	0.0616	0.0287	0.0223	0.0338	0.0262	0.0178
Acc	93.8400	97.1308	97.7686	96.6167	97.3808	98.2162
Sp	0.9384	0.9758	0.9811	0.9719	0.9845	0.9855
Ss	0.9384	0.9669	0.9743	0.9605	0.9635	0.9788
F-score	0.9384	0.9714	0.9778	0.9664	0.9741	0.9822
Cc	0.8768	0.9427	0.9554	0.9324	0.9478	0.9643

[i] Acc, accuracy; Cc, correlation coefficient; Err, error; IoU, intersection over union; Sp, specificity; Ss, sensitivity.

Comparative analysis of IoU score on training and validation images taken from autorickshaw dataset. IoU, intersection over union.

Comparative analysis of IoU score on training and validation images taken from IDD-Lite dataset. IDD-Lite, Indian driving dataset-lite; IoU, intersection over union.

Comparative analysis of IoU score on training and validation images taken from PETS dataset. IoU, intersection over union.

Comparative analysis of IoU score on training and validation images taken from MADS dataset. IoU, intersection over union; MADS, martial arts, dancing and sports.

Comparative analysis of IoU score on training and validation images taken from CT-Liver dataset. IoU, intersection over union.

Table 7:

Statistical performance analysis of LWA-MoDUNeT against other compared SOTA models

Models	Samples size = 30 p-value	Samples size = 50 p-value
UNet and LWA-MoDUNet	0.0029 × E−05	3.54 × E−09
ResNet34 and LWA-ModUNet	8.49 × E−04	7.28 × E−06
ENet and LWA-MoDUNet	3.04 × E−02	9.51 × E−03
SegNet and LWA-ModUNet	4.85 × E−02	6.39 × E−04

[i] LWA-MoDUNet, lightweight modified UNet model.

Model type 1 and type 2 errors comparison across datasets. IDD-Lite, Indian driving dataset-lite; MADS, martial arts, dancing and sports.

Table 8:

Comparative analysis of computational complexity

Model	Measures	Inference time (ms/image)
FCN-8s [28]	134.27	86.1
SegNet [29]	29.46	60.7
DeconvNet [31]	251.84	214.3
U-Net [30]	31.03	47.2
ResUNet [51]	8.10	55.8
LWA-ModUNet	5.17	36.4

References

Authors

Metrics

Articles in this issue

DOI: https://doi.org/10.2478/ijssis-2026-0014 | Journal eISSN: 1178-5608

Journal RSS Feed

Language: English

Submitted on: Jul 11, 2025

Published on: Apr 10, 2026

Published by: International Journal on Smart Sensing and Intelligent Systems

In partnership with: Paradigm Publishing Services

Publication frequency: 1 issue per year

Keywords:

semantic segmentation,

UNet,

attention module

Related subjects:

Engineering,

Introductions and overviews,

Engineering, other

© 2026 Twinkle Tiwari, Mukesh Saraswat, published by International Journal on Smart Sensing and Intelligent Systems
This work is licensed under the Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 License.

Volume 19 (2026): Issue 1 (January 2026)

Lightweight inception-UNet with attention mechanisms for semantic segmentation

Figures & Tables

Figure 1:

Figure 2:

Figure 3:

Figure 4:

Figure 5:

Figure 6:

Figure 7:

Table 1:

Figure 8:

Figure 9:

Figure 10:

Figure 11:

Figure 12:

Table 2:

Table 3:

Table 4:

Table 5:

Table 6:

Figure 13:

Figure 14:

Figure 15:

Figure 16:

Figure 17:

Table 7:

Figure 18:

Table 8:

Paradigm

My account