LMSC-UNet: A Lightweight U-Net with Modified Skip Connections for
Semantic Segmentation
Shrutika S. Sawant
1 a
, Andreas Medgyesy
1 b
, Sahana Raghunandan
1 c
and Theresa G
¨
otz
1,2 d
1
Fraunhofer Institute of Integrated Circuits, Erlangen, Germany
2
Department of Industrial Engineering and Health, Technical University of Applied Sciences Amberg-Weiden, Germany
Keywords:
Depthwise Separable Convolutions, Skip Connections, Semantic Segmentation, U-Net.
Abstract:
U-Net, an encoder-decoder architecture is the most popular choice in the semantic segmentation field due to
its ability to learn rich semantic features while handling enormous amounts of data. However, due to large
number of parameters and slow inference, deploying U-Net on devices with limited computational resources
such as mobile and embedded devices becomes challenging. To alleviate the above challenge, in this study, we
propose an efficient, lightweight, and robust encoder-decoder architecture, LMSC-UNet for semantic segmen-
tation that captures more comprehensive, contextual information and effectively learns rich semantic features.
This lightweight architecture considerably reduces the number of trainable parameters, requiring sufficiently
less memory space, training, and inference time. Skip connections in original U-Net fuse features from each
encoder block to the corresponding decoder block. This simple skip connection reduces the semantic gap to
some extent and may limit the segmentation performance. Therefore, we replace the skip connection from the
second level of U-Net with a bottleneck residual block (BRB) which helps to enhance the final segmentation
map by lessening the semantic gap between the features of decoder with the corresponding features of encoder.
Extensive experiments on various segmentation datasets from diverse domains demonstrate the effectiveness
of our proposed approach. The experimental results show that the compact model speeds up the inference
process, while still maintaining the performance. When compared to the standard U-Net, LMSC-UNet has
achieved 7× reduction in Floating Point Operations (FLOPs), and 34× reduction in model size, while main-
taining the segmentation accuracy.
1 INTRODUCTION
Semantic segmentation is an essential and fundamen-
tal task in computer vision aimed at assigning a la-
bel to each pixel in an image. This has been widely
studied in various domains, such as remote sensing
data analysis (Meng et al., 2022), medical image anal-
ysis (Al-Masni and Kim, 2021), autonomous driv-
ing (Nawaratne et al., 2020) and so on (Zhao et al.,
2024), (Zhang et al., 2021). Since the introduc-
tion of deep learning, researchers have taken inter-
est in developing semantic segmentation algorithms
based on deep neural networks (Akkus et al., 2017),
(Milletari et al., 2016). Fully Convolutional Net-
works (FCN) is the earliest deep segmentation net-
a
https://orcid.org/0000-0002-1532-947X
b
https://orcid.org/0009-0002-8129-5738
c
https://orcid.org/0009-0009-8056-3420
d
https://orcid.org/0000-0001-8751-3404
work (Long et al., 2015). Subsequently, various seg-
mentation networks were introduced based on FCN
which showed quite excellent segmentation perfor-
mance (Badrinarayanan et al., 2017), (Milletari et al.,
2016), (Chen et al., 2018). In 2015, an encoder-
decoder structure based on FCN was introduced,
called as U-Net, for segmenting biomedical images
(Ronneberger et al., 2015). By employing skip con-
nections for mapping low-resolution features to high
resolution features, U-Net produced more precise seg-
mentation maps. Due to its exceptional segmentation
ability, U-Net has gained immense popularity among
the research community and has been adopted to var-
ious research fields. Furthermore, numerous variants
of U-Net have emerged, such as, Dense-UNet (Cao
et al., 2020), UNet++ (Zhou et al., 2020), Dual-UNet
(Li et al., 2019), MSU-Net (Su et al., 2021), etc.
Despite demonstrating excellent segmentation ca-
pability and wide adoption in various domains, stan-
dard U-Net suffers from some limitations, especially
726
Sawant, S. S., Medgyesy, A., Raghunandan, S. and Götz, T.
LMSC-UNet: A Lightweight U-Net with Modified Skip Connections for Semantic Segmentation.
DOI: 10.5220/0013343800003890
Paper published under CC license (CC BY-NC-ND 4.0)
In Proceedings of the 17th International Conference on Agents and Artificial Intelligence (ICAART 2025) - Volume 2, pages 726-734
ISBN: 978-989-758-737-5; ISSN: 2184-433X
Proceedings Copyright © 2025 by SCITEPRESS – Science and Technology Publications, Lda.

when deploying it on limited resource devices for
real-time inferencing. The main challenge is that U-
Net is a large complex model encompassing multi-
ple convolutional layers with many filters and needs
an enormous number of trainable parameters. This
led to a high computational cost which demands a
large amount of memory footprint. As a result, the
standard U-Net demands almost 31 million trainable
parameters and 54 billion (G) FLOPs to get a seg-
mented output for an image size of 256 × 256. Fur-
thermore, U-Net’s encoder-decoder design does not
handle multiscale features compactly, the redundancy
in processing image scales often leads to unnecessary
computational overhead. To overcome these limita-
tions, compressing U-Net becomes essential that fa-
cilitates effective deployment in resource-constrained
environments while preserving the performance of the
original U-Net model. There exist various methods of
U-Net model compression, including model pruning
(Sawant et al., 2022a), (Sawant et al., 2022b), (Sawant
et al., 2023), quantization (Nam et al., 2024), knowl-
edge distillation (Vaze et al., 2020) and lightweight
variants of U-Net (Beheshti and Johnsson, 2020),
(El-Assiouti et al., 2023). By incorporating one
of the model compression techniques or combining
two, the U-Net can be made suitable for resource-
constrained environments without significantly com-
promising segmentation accuracy.
In this study, we propose an efficient, lightweight,
and robust encoder-decoder architecture for semantic
segmentation that captures more comprehensive con-
textual information and effectively learns rich seman-
tic features. We adopt U-Net, an encoder-decoder ar-
chitecture as a backbone. We modify the standard U-
Net architecture such that it contains less number of
trainable parameters and still yields precise segmen-
tation maps. The main contributions of this study
are summarized as: 1) The MobileNetV2 building
blocks (MBB) are substituted in the contracting path
of the standard U-Net, which significantly reduces the
number of trainable parameters. 2) The skip connec-
tion from the second level of U-Net is replaced with
a bottleneck residual block (BRB) which contains a
residual connection and a small feature extraction-
dimensionality reduction block. 3) To demonstrate
the robustness of the proposed LMSC-UNet, we eval-
uate our model on two datasets from different do-
mains.
2 PROPOSED FRAMEWORK
In this section, we discuss LMSC-UNet, a lightweight
and efficient architecture for semantic segmentation
in detail. This model adopts standard U-Net as its
baseline model, which is a fully convolutional neural
network with a symmetric encoder-decoder structure.
The proposed LMSC-UNet architecture incorporates
MBB units in the encoder of the standard U-Net (ex-
cept first layer) and the flow of features between the
second level of encoder to the corresponding decoder
is accomplished by a BRB path, as shown in Figure
1. We use the decoder of the proposed model same as
in standard U-Net. Compared to the standard U-net,
LMSC-UNet has fewer layers with lesser the num-
ber of filters in each layer. This shallow architecture
of LMSC-UNet is more suitable for deployment on
resource-constrained devices than standard U-Net due
to its faster training and real-time inference.
As mentioned above, the LMSC-UNet model fol-
lows a similar structure as U-Net with few modifica-
tions. Except for the first layer in the encoder which
has a standard convolution, the remaining layers in the
encoder of LMSC-UNet are replaced with MBB units
as shown in Figure 2. The MBB (Sandler et al., 2019)
unit is built upon the inverted residual layer which
is composed of a 1 × 1 convolution with Relu acti-
vation, a 3 × 3 depthwise separable convolution with
Relu activation, a 1 × 1 convolution with linear acti-
vation and a residual connection. The use of MBB
offers two noteworthy benefits: firstly, depthwise sep-
arable convolutions reduce the number of parameters
which in turn decreases the memory consumption and
computational cost significantly. And secondly, in-
verted residual bottleneck preserves more detailed in-
formation and facilitates efficient gradient flow dur-
ing backpropagation, improving training stability and
convergence. The use of MBB units in the second and
third layer (which are responsible for mid-level fea-
ture extractions) of LMSC-UNet, make the encoder
more efficient in capturing the contextual details re-
quired for segmentation tasks. Overall, the inclusion
of MBB units significantly reduces the computational
footprint while preserving essential features, making
the LMSC-UNet highly efficient for segmenting com-
plex scenes.
As mentioned earlier, we use the same decoder
as that of standard U-Net, which produces segmen-
tation maps at full resolution. With skip connections
at the corresponding depths of the encoder and de-
coder, deep semantic features from decoder are com-
bined with shallow, low-level features from encoder,
which results in spatially accurate and semantically
meaningful segmentation maps. Despite improving
the segmentation accuracy, simple skip connections
often show some limitations, which can hinder perfor-
mance in segmenting complex scenes. The key lim-
itation is that directly combining low-level and high-
LMSC-UNet: A Lightweight U-Net with Modified Skip Connections for Semantic Segmentation
727

Figure 1: The proposed LMSC-UNet architecture.
Figure 2: The MBB: MobileNetV2 Building Block.
Figure 3: The BRB: Bottleneck Residual Block.
level feature maps may confuse the decoder and de-
grade segmentation accuracy. To overcome this, we
used a BRB (as used in (He et al., 2016)) path in the
second level skip connection shown in Figure 1. The
BRB is a combination of basic residual block and di-
mensionality reduction block shown in Figure 3. The
first 1×1 convolution reduces the feature dimensions,
whereas the second 1 × 1 convolution expands them.
This not only reduces the redundant information, but
also focuses on most relevant features. The intermedi-
ate 3 ×3 convolution block processes the reduced fea-
tures, enriching the spatial features passed through the
skip connection. Since, this arrangement refines the
encoder feature maps before passing to the decoder,
it certainly reduces the mismatch between encoder-
decoder feature spaces. Moreover, we use BRB only
in second-level skip connection, which not only pre-
serves intermediate contextual information, but also
ICAART 2025 - 17th International Conference on Agents and Artificial Intelligence
728

avoids overloading the model with additional compu-
tations across all levels.
With a smaller number of encoder layers in
LMSC-UNet (only three), the model may capture
only fewer levels of abstraction and struggle to ac-
curately segment the tiny objects or some intricate
boundaries. To mitigate the potential loss of preci-
sion, we use a combination of Dice loss and Focal
loss.
3 EXPERIMENTAL SETUP
To demonstrate the effectiveness of the proposed
LMSC-UNet, we evaluate its performance on two
datasets for semantic segmentation from diverse do-
mains and compare with standard U-Net.
3.1 Dataset
We quantitatively analyse the segmentation perfor-
mance of the LMSC-UNet across two publicly avail-
able datasets from two different domains: 1) Biomed-
ical imaging – NuInsSeg (Mahbod et al., 2024) and
2) Radio Frequencies (RF) signal spectrum segmen-
tation dataset (Vagollari et al., 2023). These datasets
were selected because they offer diverse and com-
prehensive cases for image segmentation challenges,
which can effectively denote the robustness and gen-
eralization capabilities of the proposed model.
NuInsSeg is a dataset of nuclei in Hematoxylin
and Eosin (HE)- stained histological images, which
has 665 image patches with more than 30, 000 man-
ually segmented nuclei from 31 humans and mouse
organs. The dataset comes with binary segmentation
masks containing nuclei and non-nuclei regions.
The RF signal dataset comprises a collection of
wideband communication signals, which represent
recordings of a radio monitoring receiver captured
across a wide frequency band. It has 4000 wideband
signals that are stored using the Signal Metadata For-
mat (SigMF). To perform the task of detecting, classi-
fying, and localizing all narrowband emissions within
a wideband signal, we generated spectrogram images
of wideband signals using Short-Time Fourier Trans-
form (STFT). We analyse this dataset to perform the
task of signal detection and localization.
3.2 Implementation Details and
Evaluation Metrics
The proposed LMSC-UNet is trained and evaluated
using Nvidia Tesla V100 (32GB) GPU, Intel Xeon
Gold 6134 (”Skylake”) CPU with 16 cores, 32 threads
@3.2GHz, 96GB RAM based on Pytorch framework
(https://pytorch.org/). For model’s weight initializa-
tion, the He (also known as Kaiming) initialization
method is employed. The network parameters are op-
timized using the Adam optimizer with a batch size of
32, the initial learning rate is set to 0.01 and the learn-
ing decay rate is 0.1. Input images from all datasets
are down-sampled using linear interpolation with a
resolution of 256×256 pixels and fed into the LMSC-
UNet. Besides, models are trained for 200 epochs on
both NuInsSeg and RF signal dataset. To ensure better
performance of the model on the NuInsSeg dataset, a
5-fold cross-validation is applied to enhance the ro-
bustness of the results. For the RF signal dataset, we
split it into 70% training, 15% validation and 15%
test set. We used 133 images as validation set in
NuInsSeg dataset and 600 spectrogram images in the
test set of RF signal dataset, without data augmenta-
tion. Segmentation performance on NuInsSeg dataset
is recorded by taking an average of all folds, whereas
results on a test set are recorded for RF signal dataset.
Model’s segmentation performance is evaluated
using segmentation accuracy (ACC.), the mean
intersection-over-union (mIoU), F1 score and Dice
coefficient (DC), which are commonly used for se-
mantic segmentation tasks. To assess the efficiency
and scalability of the model, the number of param-
eters, model size and MACs (Multiply-Accumulate)
are measured as computational metrics.
4 EXPERIMENTAL ANALYSIS
We conducted a series of experiments on the proposed
LMSC-UNet for measuring its computational require-
ments and segmentation performance on above men-
tioned datasets.
4.1 Computational Results
To deploy the model on a small, embedded plat-
form for real-time inferencing, having a compact or
lightweight model is top priority. In this study, we
utilized MBB units which contain depthwise separa-
ble convolutions to reduce the number of parameters
and computations of the model. The computational
requirements of LMSC-UNet are measured in terms
of number of parameters, FLOPs (as MACs), stor-
age requirements (model size), and compared to the
LMSC-UNet without a BRB and the standard U-Net
model. LMSC-UNet without a BRB include simple
skip connections as used in U-Net. Table 1 reports
the comparison between computational requirements
LMSC-UNet: A Lightweight U-Net with Modified Skip Connections for Semantic Segmentation
729

Table 1: Computational requirements of LMSC-UNet and other models. Here, M stands for million and B stands for billion.
Model Params (M) MACs (B) Size in memory (MB) Inference time (ms/image)
Standard U-Net 31 54.76 118 6.8
LMSC-UNet (without BRB) 0.9 8.1 3.45 2.4
LMSC-UNet (Ours) 0.9 8.2 3.46 2.4
Table 2: Segmentation performance of LMSC-UNet and other models on NuInsSeg dataset.
Model ACC.(%) mIoU F1 DC
Standard U-Net 93.57 0.6817 0.8123 0.8094
LMSC-UNet (without BRB) 94.62 0.7102 0.8335 0.8301
LMSC-UNet (Ours) 94.60 0.7123 0.8346 0.8313
Figure 4: Learning curves. (a) Standard U-Net (b) LMSC-UNet without BRB (c) LMSC-UNet (Ours).
of LMSC-UNet and other models. By incorporat-
ing MBB units in the encoder of U-Net, the num-
ber of parameters of our model is only 0.9M and the
model size is only 3.45MB, which is not even a quar-
ter of that of standard U-Net. Meanwhile, LMSC-
UNet needs approximately 8.1B MACs, which are 7×
fewer compared to the standard U-Net. After adopt-
ing a BRB, a bottleneck block in the second level of
skip connection, which adds a small computational
overhead, the number of parameters and MACs uti-
lization of LMSC-UNet are slightly higher than that
of LMSC-UNet without a BRB (unaltered skip con-
nection). Furthermore, LMSC-UNet with and with-
out BRB is almost three times faster in inference than
that of standard U-Net.
4.2 Segmentation performance on
NuInsSeg Dataset
To verify the improvement in the performance due to
adoption of MBB units and BRB in U-Net, the seg-
mentation results of U-Net, LMSC-UNet (without a
BRB) and the proposed LMSC-UNet are presented in
Table 2. Both LMSC-UNet without BRB and with
BRB are seen to achieve comparable performance in
terms of segmentation accuracy, mIoU, F1 score and
DC, while being computationally more efficient than
the standard U-Net model. More specifically, BRB
contributes to the higher segmentation performance
of LMSC-UNet in comparison to the standard U-Net.
Compared to the standard U-Net, when standard con-
volution blocks in the encoder are replaced with MBB
units, the ACC., mIoU and DC of our lightweight
model are enhanced by almost 1%, 2.85% and 2%,
respectively. By adopting BRB in the second level of
skip connection, the segmentation performance is fur-
ther improved, which clearly demonstrates the bene-
fits of refining intermediate features before combining
with decoder feature maps. Since we have retained
unaltered skip connections at other levels, the decoder
gets sufficient spatial details and avoids possible accu-
racy degradation.
For a more comprehensive validation, especially
to validate the effect of MBB units and BRB path,
the learning curves are also provided in Figure 4. As
shown in Figure 4(a), the training process of stan-
dard U-Net is unstable, whereas LMSC-UNet with-
out BRB improves the stability of training, however,
suffers from little overfitting. The reason behind the
overfitting is, MBB units mainly prioritize the param-
ICAART 2025 - 17th International Conference on Agents and Artificial Intelligence
730

Figure 5: Qualitative results on NuInsSeg dataset. (a) Input image (b) Ground truth (c) Standard U-Net (d) LMSC-UNet
without BRB (e) LMSC-UNet (Ours).
Table 3: Ablation study of BRB adopted at different levels of skip connections in LMSC-UNet.
Model Params (M) MACs (B) ACC.(%) mIoU
LMSC-UNet-1 0.9 8.2 94.56 0.7093
LMSC-UNet-2 (Ours) 0.91 8.2 94.60 0.7123
LMSC-UNet-3 0.93 8.25 94.61 0.7079
LMSC-UNet-12 0.91 8.4 94.39 0.7022
LMSC-UNet-23 0.94 8.4 94.80 0.7071
LMSC-UNet-123 0.95 8.5 94.79 0.7141
eter and computation reductions. However, this strat-
egy may lead to loss of some fine details which af-
fects the segmentation performance. With inclusion
of BRB, the proposed model passes only mid-level
relevant features to the decoder, which further reduces
overfitting and follows a stable training process. The
BRB unit not only ensures smooth gradient propaga-
tion but also avoids vanishing gradients, leading to
better training convergence. Qualitative results on the
NuInsSeg dataset are displayed in Figure 5. Red rect-
angles in the Figure represent the regions with precise
and accurate segmentation. Due to the large number
of parameters, the standard U-Net may have memo-
rized training data, including noise which resulted in
wrong predictions. LMSC-UNet without BRB also
produced wrong predictions. On the other hand, the
BRB unit in the proposed LMSC-UNet passed only
meaningful features and produced segmentation maps
which align more closely with the ground truth, as
shown in Figure 5(e).
Overall, the combination of MBB and BRB helps
the model to extract rich features and ensures that
only meaningful and less redundant features are
passed to the decoder, enhancing the model’s segmen-
tation ability.
LMSC-UNet: A Lightweight U-Net with Modified Skip Connections for Semantic Segmentation
731

Figure 6: Visualization of feature maps. (a) LMSC-UNet-1 (b) LMSC-UNet-2 (Ours) (c) LMSC-UNet-3 (d) LMSC-UNet-12
(e) LMSC-UNet-23 (f) LMSC-UNet-123.
Table 4: Segmentation performance of LMSC-UNet and other models on RF signal dataset.
Model ACC.(%) mIoU F1 DC
Standard U-Net 98.19 0.9172 0.9547 0.9545
LMSC-UNet (without BRB) 97.29 0.8904 0.9394 0.9392
LMSC-UNet (Ours) 97.27 0.8903 0.9391 0.9390
4.3 Ablation Study
In order to find the level at which BRB can be in-
corporated in the skip connection that ensures only
relevant features will be passed to the decoder, we
conducted a series of ablation experiments on inclu-
sion of BRB at various levels of skip connections. Ta-
ble 3 presents the segmentation performance and cor-
responding computational demands of different ver-
sions of the proposed model. The first level skip con-
nection in the LMSC-UNet is responsible for low-
level, fine grained spatial information, second level
carries intermediate feature representations, whereas
third level carries high-level abstract features to the
corresponding decoder. Therefore, refining low-level
features may lose fine spatial details which are critical
for reconstructing complex object structures, whereas
refining high-level abstract features may not con-
tribute much to the final segmentation maps. This has
been confirmed by the results recorded in Table 3. In
comparison to BRB at other levels of skip connec-
tions, the proposed model (BRB at second level) was
able to achieve excellent segmentation performance
with minimal number of parameters and MACs. Ad-
ditionally, feature maps extracted from the last de-
coder block corresponding to each modification is
shown in Figure 6. Since the last decoder block
plays a critical role in producing the final segmenta-
tion map, the feature maps from this layer highlights
the richness of semantic details available for segmen-
ICAART 2025 - 17th International Conference on Agents and Artificial Intelligence
732

Figure 7: Qualitative results on RF Signal dataset. (a) A spectrogram image 1 (b) A spectrogram image 2.
tation. It is clearly seen that, BRB at the second level
of skip connection (our model) passes more relevant
features and produces more accurate segmentation.
4.4 Segmentation performance on RF
Signal Dataset
To further demonstrate the robustness and general-
izability of the proposed model, we evaluate it on
RF signal dataset, where input data vary in format,
pattern, object orientations than that of histological
imaging dataset, NuInsSeg. These experimentations
also ensure that the LMSC-UNet is reliable and ver-
satile. Table 4 presents segmentation performance
of the proposed model and other models on RF sig-
nal dataset. The dataset is quite larger than that of
the NuInsSeg dataset. Therefore, it is quite inter-
esting to analyse our lightweight model with a large
dataset as well. As shown in Table 4, the segmen-
tation results of LMSC-UNet with and without BRB
are marginally lower than that of standard U-Net.
The large number of parameters contributes to higher
performance of standard U-Net apparently, whereas
our lightweight model slightly struggles to maintain
the segmentation performance. Albeit showing lower
segmentation performance than standard U-Net, our
model with and without BRB shows a better trade-
off by reaching a mIoU of 0.8903 while requiring less
than 1M parameters. These results highlight the gen-
eralizability of our lightweight model when dealing
with relatively large datasets. Qualitative results on
RF signal dataset are displayed in Figure 7. A spectro-
gram image of a wideband signal containing four sig-
nals of long duration, its ground truth and segmented
maps generated by different models are shown in Fig-
ure 7(a), whereas Figure 7(b) shows a spectrogram
image of a wideband signal containing four signals of
short duration, its ground truth and segmented maps
generated by different models. The segmented maps
generated by standard U-Net, LMSC-UNet with and
without BRB look almost similar. Despite achiev-
ing lowest segmentation performance on test data, our
lightweight LMSC-UNet model accurately detected
even signals with short duration.
5 CONCLUSION
In this study, we proposed LMSC-UNet, a lightweight
and robust architecture for accurate and precise seg-
mentation. The incorporated MBB units in the en-
coder not only reduces the number of trainable pa-
rameters but also facilitates faster inference by using
fewer MACs. Additionally, using BRB in the sec-
ond level of skip connection, the proposed model en-
sured that only relevant features are passed to the de-
coder which further enhanced the segmentation ac-
curacies. We evaluated our LMSC-UNet model on
two different datasets from different domains and ob-
served that it obtains comparable segmentation per-
formance by utilizing a smaller number of parameters
(∼ 0.9M), significantly fewer MACs and low mem-
ory consumption. Overall, the proposed LMSC-UNet
model obtained a better trade-off between computa-
tional resources and segmentation performance, mak-
ing it suitable for limited-resource devices and real-
time applications.
LMSC-UNet: A Lightweight U-Net with Modified Skip Connections for Semantic Segmentation
733

ACKNOWLEDGEMENTS
This work was funded by FH-KOOP funding Project
of Weiden Erlangen Cooperation for Sparse AI in
Life Sensing under internal grant from Fraunhofer
gesellschaft and supported by Fraunhofer Institute for
Integrated Circuits (IIS) by providing infrastructure to
carry out the research work.
REFERENCES
Akkus, Z., Galimzianova, A., Hoogi, A., Rubin, D. L., and
Erickson, B. J. (2017). Deep learning for brain MRI
segmentation: state of the art and future directions. J
Digit Imag, 30(4):449–459.
Al-Masni, M. A. and Kim, D.-H. (2021). Cmm-net: Con-
textual multi-scale multi-level network for efficient
biomedical image segmentation. Scientific Reports,
11(1):1–18.
Badrinarayanan, V., Kendall, A., and Cipolla, R. (2017).
Segnet: A deep convolutional encoder-decoder ar-
chitecture for image segmentation. IEEE Transac-
tions on Pattern Analysis and Machine Intelligence,
39(12):2481–2495.
Beheshti, N. and Johnsson, L. (2020). Squeeze u-net: A
memory and energy efficient image segmentation net-
work. In 2020 IEEE/CVF Conference on Computer
Vision and Pattern Recognition Workshops (CVPRW),
pages 1495–1504.
Cao, Y., Liu, S., Peng, Y., and Li, J. (2020). Denseunet:
densely connected unet for electron microscopy image
segmentation. IET Image Processing, 14(12):2682–
2689.
Chen, L.-C., Papandreou, G., Kokkinos, I., Murphy, K., and
Yuille, A. L. (2018). Deeplab: Semantic image seg-
mentation with deep convolutional nets, atrous con-
volution, and fully connected crfs. IEEE Transac-
tions on Pattern Analysis and Machine Intelligence,
40(4):834–848.
El-Assiouti, H. S., El-Saadawy, H., Al-Berry, M. N., and
Tolba, M. F. (2023). Lite-srgan and lite-unet: Toward
fast and accurate image super-resolution, segmenta-
tion, and localization for plant leaf diseases. IEEE
Access, 11:67498–67517.
He, K., Zhang, X., Ren, S., and Sun, J. (2016). Deep resid-
ual learning for image recognition. In 2016 IEEE Con-
ference on Computer Vision and Pattern Recognition
(CVPR), pages 770–778.
Li, X., Wang, Y., Tang, Q., Fan, Z., and Wu, J. (2019). Dual
U-Net for the segmentation of overlapping glioma nu-
clei. IEEE Access, 7:84040–84052.
Long, J., Shelhamer, E., and Darrell, T. (2015). Fully con-
volutional networks for semantic segmentation. In
IEEE Conf Comp Vision and Pattern Recog, pages
3431–3440. IEEE.
Mahbod, A., Polak, C., Feldmann, K., Khan, R., Gelles, K.,
Dorffner, G., Woitek, R., Hatamikia, S., and Ellinger,
I. (2024). Nuinsseg: A fully annotated dataset for nu-
clei instance segmentation in h&e-stained histological
images. Scientific Data, 11(1):1–7.
Meng, X., Yang, Y., Wang, L., Wang, T., Li, R., and Zhang,
C. (2022). Class-guided swin transformer for seman-
tic segmentation of remote sensing imagery. IEEE
Geoscience and Remote Sensing Letters, 19:1–5.
Milletari, F., Navab, N., and Ahmadi, S.-A. (2016). V-
Net: Fully convolutional neural networks for volumet-
ric medical image segmentation. In Fourth IEEE Int
Conf 3D Vision, pages 565–571, Stanford, CA, USA.
IEEE.
Nam, M., Oh, S., and Lee, J. (2024). Quantization of u-
net model for self-driving. In 2024 10th International
Conference on Applied System Innovation (ICASI),
pages 1–3.
Nawaratne, R., Alahakoon, D., De Silva, D., and Yu, X.
(2020). Spatiotemporal anomaly detection using deep
learning for real-time video surveillance. IEEE Trans-
actions on Industrial Informatics, 16(1):393–402.
Ronneberger, O., Fischer, P., and Brox, T. (2015). MIC-
CAI2015, chapter U-Net: Convolutional networks for
biomedical image segmentation. Springer.
Sandler, M., Howard, A., Zhu, M., Zhmoginov, A., and
Chen, L.-C. (2019). Mobilenetv2: Inverted residuals
and linear bottlenecks.
Sawant, S. S., Bauer, J., Erick, F. X., Ingaleshwar, S.,
Holzer, N., Ramming, A., Lang, E. W., and G
¨
otz,
T. (2022a). An optimal-score-based filter pruning for
deep convolutional neural networks. Applied Intelli-
gence, 52:17557–17579.
Sawant, S. S., Erick, F. X., G
¨
ob, S., Holzer, N., Lang,
E. W., and G
¨
otz, T. (2023). An adaptive binary par-
ticle swarm optimization for solving multi-objective
convolutional filter pruning problem. Journal of Su-
percomputing, 79:13287–13306.
Sawant, S. S., Wiedmann, M., G
¨
ob, S., Holzer, N., Lang,
E. W., and G
¨
otz, T. (2022b). Compression of
deep convolutional neural network using additional
importance-weight-based filter pruning approach. Ap-
plied Sciences, 12(21).
Su, R., Zhang, D., Liu, J., and Cheng, C. (2021). MSU-Net:
Multi-scale U-Net for 2D medical image segmenta-
tion. Frontiers in Genetics, 12:639930.
Vagollari, A., Hirschbeck, M., and Gerstacker, W. (2023).
An end-to-end deep learning framework for wideband
signal recognition. IEEE Access, 11:52899–52922.
Vaze, S., Xie, W., and Namburete, A. I. L. (2020). Low-
memory cnns enabling real-time ultrasound segmen-
tation towards mobile deployment. IEEE Journal
of Biomedical and Health Informatics, 24(4):1059–
1069.
Zhang, J., Zhu, H., Wang, P., and Ling, X. (2021). Att
squeeze u-net: A lightweight network for forest fire
detection and recognition. IEEE Access, 9:10858–
10870.
Zhao, P., Li, Z., You, Z., Chen, Z., Huang, T., Guo, K., and
Li, D. (2024). Se-u-lite: Milling tool wear segmenta-
tion based on lightweight u-net model with squeeze-
and-excitation module. IEEE Transactions on Instru-
mentation and Measurement, 73:1–8.
Zhou, Z., Siddiquee, M. M. R., Tajbakhsh, N., and Liang, J.
(2020). Unet++: redesigning skip connections to ex-
ploit multiscale features in image segmentation. IEEE
Trans Med Imaging, 39:1856–1867.
ICAART 2025 - 17th International Conference on Agents and Artificial Intelligence
734