Recognition-Oriented Low-Light Image Enhancement Based on
Global and Pixelwise Optimization
Seitaro Ono
1
, Yuka Ogino
2
, Takahiro Toizumi
2
, Atsushi Ito
2
and Masato Tsukada
1
1
University of Tsukuba, Ibaraki, Japan
2
NEC Corporation, Kanagawa, Japan
Keywords: Low-Light Image Enhancement, Image Recognition.
Abstract: In this paper, we propose a novel low-light image enhancement method aimed at improving the performance
of recognition models. Despite recent advances in deep learning, the recognition of images under low-light
conditions remains a challenge. Although existing low-light image enhancement methods have been
developed to improve image visibility for human vision, they do not specifically focus on enhancing
recognition model performance. Our proposed low-light image enhancement method consists of two key
modules: the Global Enhance Module, which adjusts the overall brightness and color balance of the input
image, and the Pixelwise Adjustment Module, which refines image features at the pixel level. These modules
are trained to enhance input images to improve downstream recognition model performance effectively.
Notably, the proposed method can be applied as a frontend filter to improve low-light recognition performance
without requiring retraining of downstream recognition models. Experimental results demonstrate that our
method improves the performance of pretrained recognition models under low-light conditions and its
effectiveness.
1 INTRODUCTION
In recent years, deep learning-based image
recognition models have achieved significant
advancements (Wang et al., 2020; Muhammad et al.,
2022; Zheng et al., 2023), demonstrating
progressively improved performance. However, these
models highly rely on large datasets (Geiger et al.,
2012; Andriluka et al., 2014; Lin et al., 2014; Güler
et al., 2018) consisting of high-quality images
captured under ideal good lighting conditions,
presenting substantial challenges when applied to
low-light environments (Tian et al., 2023; Ono et al.,
2024). In low-light settings, factors such as low
exposure and reduced contrast severely impair image
visibility. To improve the visibility of low-light
images, deep learning-based low-light image
enhancement methods have been developed (Wang et
al., 2020; Li et al., 2021; Liu et al., 2021; Tian et al.,
2023). However, these methods are primarily
designed for human vision, without consideration for
the performance of recognition models. Consequently,
conventional enhancement methods may cause
excessive smoothing or amplifying noise in output
images, potentially degrading recognition model
performance (Ono et al., 2024; Ogino et al., 2024).
To address the reduction in recognition
performance under low-light conditions, we propose
a novel low-light image enhancement method that
converts input images into more recognizable images
for downstream recognition models. The proposed
enhancement method comprises two modules: the
Global Enhancement Module (GEM), which adjusts
the overall brightness of input images, and the
Pixelwise Adjustment Module (PAM), which refines
image features on a pixel level. GEM adjusts the
brightness of the entire input image, producing a
globally enhanced image, 𝐼
. PAM, on the other
hand, estimates an optimal pixel-wise correction map,
𝑓
, that records adjustment values for each pixel in
𝐼
. By taking the linear combination of 𝐼
and 𝑓
, we generate the output image 𝐼
.
Our enhancement model is trained to minimize a
model specific loss function for a downstream
recognition model by optimizing the output image
𝐼
to make it more suitable for the downstream
recognition model. By optimizing our low-light
enhancement method to reduce the model specific
loss for the downstream recognition model, we
222
Ono, S., Ogino, Y., Toizumi, T., Ito, A. and Tsukada, M.
Recognition-Oriented Low-Light Image Enhancement Based on Global and Pixelwise Optimization.
DOI: 10.5220/0013316800003912
Paper published under CC license (CC BY-NC-ND 4.0)
In Proceedings of the 20th International Joint Conference on Computer Vision, Imaging and Computer Graphics Theory and Applications (VISIGRAPP 2025) - Volume 3: VISAPP, pages
222-232
ISBN: 978-989-758-728-3; ISSN: 2184-4321
Proceedings Copyright © 2025 by SCITEPRESS – Science and Technology Publications, Lda.
enhance the input image in a way that facilitates
improved recognition performance.
To evaluate the effectiveness of our proposed
method under low-light conditions and its
generalizability across different recognition tasks, we
conducted experiments on two tasks: single-person
pose estimation and semantic segmentation. The
experimental results demonstrate that our proposed
method effectively enhances the performance of
pretrained recognition models under low-light
conditions and improves their performance across
different recognition tasks, validating its
effectiveness and generalizability.
2 RELATED WORKS
Image recognition under low-light conditions is a
critical challenge in computer vision (Liang et al.,
2021; Wang et al., 2022), and various methods have
been proposed to improve image visibility in such
environments. In recent years, various low-light image
enhancement techniques based on convolutional neural
networks have emerged (Wang et al., 2020; Li et al.,
2021; Tian et al., 2023). Zero-DCE (Guo et al., 2020)
formulates low-light image enhancement as an image-
specific tone curve estimation task in a deep network.
By designing a non-reference loss function, Zero-DCE
can be effectively trained without paired datasets and
has been shown to perform well across diverse lighting
conditions.
LLFLow (Wang et al., 2022) is a supervised low-
light image enhancement framework based on
normalizing flow, designed to model the distribution
of well-exposed images accurately. By utilizing this
approach, LLFLow achieves enhanced structural
detail preservation across varied contexts, leading to
superior restoration quality in low-light conditions.
Although these conventional methods are
effective for enhancing low-light images, many of
these data-driven methods rely heavily on large-
paired data sets of low-light and bright images (Cai et
al., 2018; Chen et al., 2018; Wei et al., 2018). Not
only is the collection of such data sets very costly, but
this dependence also limits their practicality by
limiting the situations in which they can be applied.
Furthermore, these methods are mainly designed to
improve image visibility for human vision and often do
not consider their impact on downstream recognition
models. Consequently, applying these techniques to
low-light images can unintentionally discard features
vital for the downstream recognition model during the
enhancement process (Ono et al., 2024).
The Image-Adaptive Learnable Module (IALM)
(Ono et al., 2024) is a novel low-light image
enhancement method explicitly designed to improve
downstream recognition model performance rather
than human perceptibility. Unlike conventional
enhancement methods, which minimize a distance
function between restored images and ground-truth
images, IALM maximizes recognition performance by
optimizing image correction applied to input images so
as to minimize the loss function specific to the
downstream recognition model. This approach ensures
that the enhancement process adapts input images to
improve recognition model performance, effectively
enhancing features essential for recognition.
IALM consists of three image processing
modules: the Exposure Module, which adaptively
adjusts the exposure of an input image to suit the
downstream recognition model; the Gamma Module,
which adaptively adjusts the image contrast; and the
Smoothing Module, which performs noise reduction.
Each of these modules includes a parameter predictor
that estimates the image processing parameters
needed for each adjustment. Although IALM utilizes
a relatively simple image processing module, it has
been shown to improve the performance of
downstream recognition models compared to
conventional low-light image enhancement methods
(Ono et al., 2024). However, IALM only supports
global image adjustment, which imposes a limitation
in that it cannot perform detailed pixel-level
correction for low-light images. Consequently, to
further enhance downstream recognition model
performance, it is necessary to develop low-light
image enhancement methods tailored to recognition
models, equipped with more fine-grained image
processing capabilities.
3 PROPOSED METHOD
IALM performs exposure correction, gamma
adjustment, and noise
reduction on input low-light
images in accordance with the characteristics of a
downstream recognition model. By uniformly
adjusting the entire image, IALM enhances
recognition model performance. We think that
applying pixelwise feature adjustment may further
improve the performance of the recognition model. In
this study, we propose a low-light image
enhancement method designed to improve
recognition performance better, comprising two
modules: the Global Enhancement Module (GEM),
which globally adjusts the brightness and the color
balance of an input image, and the Pixelwise
Recognition-Oriented Low-Light Image Enhancement Based on Global and Pixelwise Optimization
223
Figure 1: The overall training pipeline for the proposed Low-Light Image Enhancement method. Input low-light image is fed
to GEM for global correction of exposure and color balance. The corrected image is then fed into the PAM, which corrects
the image at the pixel level, yielding the final output image. The output image is then input to the downstream recognition
model (Frozen Weights), where model specific losses are calculated. The weights of the proposed low-light image
enhancement model are updated to minimize model-specific losses. This process corrects the images to improve the
performance of the downstream recognition model.
Adjustment Module (PAM), which performs
pixelwise corrections to enhance image features. An
overview of our proposed framework is shown in
Figure 1.
An input image 𝐼
is first downsampled using
bilinear interpolation, yielding a low-resolution
image 𝐼
. 𝐼
is then fed into GEM, which
determines the optimal correction parameters to
adjust the 𝐼
linearly. While downsampling
reduces high-frequency components of images, it
allows GEM to focus on capturing global information,
such as exposure. By applying GEM’s predicted
correction parameters to the 𝐼
, we obtain a
globally enhanced image 𝐼
. 𝐼
is then
passed to PAM, which creates a pixelwise adjustment
map, 𝑓
, that refines image features beneficial to a
downstream recognition model. The correction map
𝑓
is added to 𝐼
, resulting in the output image
𝐼
. The output image,
𝐼
, is fed into the
recognition model, where a model-specific loss is
calculated based on the difference between the
model’s predictions and the ground truth. After
calculating this loss, the low-light image
enhancement model is trained to minimize it, thereby
improving image features in a way that enhances
recognition performance.
Since our method focuses on optimizing the
image enhancement method as a pre-processing step
for the recognition model, no gradient updates are
applied to the recognition model during training. This
training strategy leverages the pretrained knowledge
of the recognition model without requiring its
retraining. It effectively highlights image features
beneficial to the recognition model and boosts
recognition performance. Our low-light enhancement
method is implemented with a compact convolutional
neural network architecture containing only 577k
parameters. Compared to conventional low-light
enhancement methods, which often require millions
of parameters and multiple convolutional layers, our
method is lightweight, highly practical, and suitable
for a wide range of real-world applications.
3.1 Global Enhance Module
Global Enhancement Module (GEM) adaptively
adjusts the global brightness and color balance of an
input image. GEM is a lightweight convolutional
neural network composed of six convolutional layers.
Given that processing high-resolution images with
convolutional neural networks requires substantial
computational resources, GEM is fed a low-
resolution image, 𝐼
, created by downsampling the
input image 𝐼
to a 32×32 resolution using
bilinear interpolation to reduce the resource demands.
Although the downsampling process removes
high-frequency features from 𝐼
, the primary goal
of GEM is to understand global information such as
exposure and calculate adaptive exposure correction
parameters for 𝐼
; thus, a low-resolution input is
sufficient and appropriate for this purpose.
The low-resolution image 𝐼
is fed into GEM,
which then predicts three correction coefficients
𝑎
,𝑎
,𝑎
by a convolutional neural network for
adjusting the exposure and the color balance of 𝐼
. For each pixel 𝑖 in 𝐼
, with initial pixel values
𝑟
, 𝑔
, 𝑏
and adjusted pixel values 𝑟
, 𝑔
,𝑏
,
the following operation is performed.
𝑟
,𝑔
,𝑏
=
𝑎
𝑟
, 𝑎
𝑔
, 𝑎
𝑏
1
VISAPP 2025 - 20th International Conference on Computer Vision Theory and Applications
224
This correction yields the exposure-enhanced and
color-balanced image, 𝐼
.
3.2 Pixelwise Adjustment Module
Pixelwise Adjustment Module (PAM) is designed to
predict a pixelwise adjustment map that enhances
beneficial image features to improve the performance
of downstream recognition models.
To obtain features beneficial for downstream
recognition models at the pixel level, PAM utilizes a
tiny fully convolutional neural network based on the
UNet (Ronneberger et al, 2015) architecture. This
fully convolutional network has a symmetric encoder
and decoder structure.
The encoder part consists of four convolutional
blocks. Each convolutional block of the encoder
consists of two convolutional layers, a batch
normalization layer, and a ReLU activation function
to extract important information while reducing the
resolution of the feature map.
The decoder section consists of four transposed
convolutional blocks. Each transposed convolution
block of the decoder section consists of a transposed
convolution layer, a convolution layer, a batch-
normalization layer, and a ReLU activation function.
The UNet architecture allows fine pixel
information to be refined while gradually increasing
the resolution of the feature map obtained from the
encoder at the decoder. Each transposed
convolutional block combines features from the
corresponding encoder layer via skip connections.
Finally, we obtain a pixelwise adjustment map of the
same size as the input image size.
In the image processing workflow of PAM, the
exposure-corrected and color-balanced image 𝐼
,
produced by GEM is first fed into PAM. The PAM
then predicts a pixelwise adjustment map of the same
resolution as 𝐼
. The predicted pixelwise
adjustment map, 𝑓
, is added to 𝐼
to generate
the output image 𝐼
, as described by the following
operation:
𝐼
=𝐼
+𝑓
2
This process produces an output image optimized
to improve performance in downstream recognition
models.
4 EXPERIMENTS
We demonstrated the effectiveness of our proposed
method under low-light conditions. We quantitatively
evaluated the performance of the proposed method in
low-light conditions. We conducted experiments
across different tasks to demonstrate the
generalizability of the proposed method. We also
compared the enhanced images with those processed
by conventional methods for a qualitative evaluation.
We also conducted an ablation study to evaluate the
effectiveness of each of the two proposed modules.
4.1 Implementation Details
We evaluated the effectiveness and generalizability
of our proposed method using two different
recognition tasks: single person pose estimation and
semantic segmentation.
Single person pose estimation involves predicting
the coordinates of keypoints, such as the head,
shoulders, and elbows, for a single person present in
the input image (Zheng et al., 2023). On the other
hand, semantic segmentation aims to predict pixel-
wise classifications for various class categories
present in the image (Garcia-Garcia et al., 2018).
For the single person pose estimation task, we
used the pose estimation model proposed by Lee et al.
(Lee et al., 2023) as our recognition model. This
model is pre-trained on the ExLPose dataset (Lee et
al., 2023), specifically designed for pose estimation
in low-light conditions, enabling it to perform pose
estimation in both dark and well-lit environments.
We evaluated the performance of Lee et al.'s pose
estimation model both with and without applying our
proposed LLIE method.
In the semantic segmentation task, we utilized
DeepLabV3+ (Chen et al., 2018) as a recognition
model. DeepLabV3+ has been pre-trained on the
Cityscapes dataset (Cordts et al., 2016), which
consists of semantic segmentation data collected from
daytime urban street scenes. Unlike the model by Lee
et al., DeepLabV3+ is not designed for estimation in
low-light conditions. To evaluate the effectiveness of
our proposed method under low-light conditions, we
used the NightCity dataset (Tan et al., 2021), which
comprises nighttime city driving scenes, as test data
in the semantic segmentation task. Similar to the
single person pose estimation task, we evaluated our
method's effectiveness by comparing the
performance of the pre-trained DeepLabV3+ on low-
light test images from NightCity with and without our
method.
During the training of our proposed method, we
froze the weights of the downstream recognition
model in both recognition tasks and focused solely on
training the low-light image enhancement method. To
evaluate the effectiveness of our method against
Recognition-Oriented Low-Light Image Enhancement Based on Global and Pixelwise Optimization
225
conventional methods, we compared it with Zero-
DCE, LLFLow, and IALM. For the single person
pose estimation task, we trained these methods using
images from the ExLPose dataset. In the Semantic
Segmentation task, we trained each low-light image
enhancement method on the training data from the
NightCity dataset. Since the NightCity dataset does
not contain pairs of bright images corresponding to
low-light images, LLFLow, which is a supervised
method, cannot be trained. So, we used LLFLow
trained on the LOL dataset (Wei et al., 2018) in this
experiment. The LOL dataset is a widely used
benchmark dataset for low-light image enhancement,
containing 500 pairs of low-light and bright images
primarily captured indoors.
During the training of our proposed method, we
adopted the Adam Optimizer. We set the learning rate
to 5×10
and the batch size to 8. We conducted all
experiments using PyTorch, with training performed
on an RTX 3060.
4.2 Datasets
4.2.1 ExLPose
The ExLPose dataset is designed for human pose
estimation under low-light conditions. It provides
pairs of bright and low-light images. The dataset
includes 2,065 training pairs and 491 test pairs, along
with annotations for bounding boxes and keypoint
coordinates of the individuals in the images. The test
data is categorized into four subsets based on mean
pixel intensity: LL-N (Normal), LL-H (Hard), LL-E
(Extreme), and LL-A (All). The mean pixel intensity
for each subset is 3.2, 1.4, 0.9, and 2.0, respectively.
In this study, we utilized the training data for pre-
training the recognition model and training all low-
light image enhancement methods. During the
evaluation experiments, we conducted assessments
on all subsets of the ExLPose’s test data.
4.2.2 Cityscapes
The Cityscapes dataset consists of images of daytime
urban street scenes for semantic segmentation. It
serves as a benchmark dataset for segmentation tasks.
The dataset includes 2,975 training images, 500
validation images, and 1,525 test images, along with
pixel-wise annotations for 19 categories. Each image
has a resolution of 2048 × 1024 pixels. In this study,
we utilized only the training images for pre-training
DeepLabV3+.
4.2.3 NightCity
The NightCity dataset consists of images of nighttime
urban street scenes, designed for semantic
segmentation tasks. This dataset provides 2,998
training images with a size of 1024 × 512 pixels and
1,299 test images. It also offers pixelwise annotations
for the same 19 categories as the Cityscapes dataset.
In this study, we utilized this dataset for both training
and testing all low-light image enhancement methods.
4.3 Evaluation Protocol
For the single person pose estimation task, we
adopted the Average Precision (AP) score based on
object keypoint similarity (OKS) as our evaluation
metric. We reported the mean AP over OKS
thresholds ranging from 0.5 to 0.95, with a step size
of 0.05. The AP value increases with higher accuracy
in keypoint estimation.
For the semantic segmentation task, we adopted
the mean of category-wise intersection-over-union
(mIoU) as the evaluation metric. The mIoU value
increases with higher accuracy in segmentation,
indicating better recognition performance.
4.4 Experimental Results
First, we present the experimental results for the
single person pose estimation task in Table 1. The
results show that applying conventional low-light
image enhancement methods, such as Zero-DCE and
LLFLow, which do not account for the characteristics
of downstream recognition models, led to decreased
recognition performance compared to when no
enhancement was applied.
In contrast, IALM, which was trained to
maximize the prediction performance of the
downstream recognition model, successfully
improved the performance across all subsets of the
test data. Due to the limitations of IALM's correction
capabilities, the improvement in performance was
marginal.
On the other hand, our proposed method achieved
even higher performance on all subsets of the
ExLPose’s test data compared to conventional
methods, confirming its effectiveness.
We also present in Table 2 the number of parameters
for each method as well as the latency required to process
a single input image on an RTX3060 GPU. The input
images in the single person pose estimation task were 256
× 192-pixel RGB images.
VISAPP 2025 - 20th International Conference on Computer Vision Theory and Applications
226
Figure 2: Qualitative evaluation results of the single person pose estimation task. Conventional low-light image enhancement
methods can increase image visibility, but do not necessarily improve the performance of the downstream recognition model.
The output images obtained with our method has purple noise, which is not high quality for human vision. However, it allows
for highly accurate pose estimation.
Table 1: Quantitative evaluation results of the single person
pose estimation task.
AP@0.5-0.95 ↑
Model LL-N LL-H LL-E LL-A
Lee et al. 42.1 33.8 18.0 32.4
LLFLow and
Lee et al.
9.3 4.6 1.2 5.4
Zero-DCE and
Lee et al.
34.7 28.6 16.7 27.5
IALM and
Lee et al.
42.6 34.1 20.0 33.2
Ours and
Lee et al.
43.6 34.4 21.0 34.1
According to Table 2, the proposed method
demonstrates significantly faster performance than
other competing methods. Our method also has
relatively fewer parameters than conventional
methods that require many convolutional layers, such
as LLFLow. Therefore, it can effectively improve the
Table 2: Comparison of the processing speed of each method
in the single person pose estimation task. Input images are
RGB images of size 256 × 192-pixel in this task.
Model Params [M] Latency [ms]
Lee et al. 27.4 13.8
LLFLow and
Lee et al.
66.3 339.8
Zero-DCE and
Lee et al.
27.5 16.5
IALM and
Lee et al.
27.9 18.7
Ours and
Lee et al.
28.0 16.4
performance of existing trained recognition models in
low-light conditions at a low computational cost.
Figure 2 illustrates the qualitative evaluation
results for the single person pose estimation task. The
experimental results indicated that LLFlow
demonstrates excessive smoothing in the output
images, while Zero-DCE introduces additional noise,
Recognition-Oriented Low-Light Image Enhancement Based on Global and Pixelwise Optimization
227
which may contribute to recognition errors. IALM
successfully adapted the exposure to align with the
characteristics of the downstream recognition model
and managed to suppress noise. However, some cases
of recognition failure still occurred. Our proposed
method generated purple artifacts at the top and
bottom of the output images, which may not be
visually pleasing for human vision. Nonetheless, it
achieved higher recognition performance compared
to conventional methods.
The experimental results for the semantic
segmentation task are shown in Table 3. Similar to the
previous task, conventional low-light image
enhancement methods Zero-DCE and LLFLow,
which do not take into account the characteristics of
the downstream recognition model, led to a reduction
in performance. In contrast, IALM achieved an
improvement in performance, although the
improvement was marginal.
Our proposed method successfully improved the
performance of DeepLabV3+, which was trained
solely on images captured during the day, achieving
approximately a 1.87-fold improvement under low-
light conditions. We also show in Table 4 the number
of parameters for each method and the latency
required to process one input image on the RTX3060
GPU for the semantic segmentation task. The input
images for the semantic segmentation task were 1024
× 512-pixel RGB images. As shown in Table 4, the
proposed method exhibited superior speed compared
to other competitive methods, even when handling
large 1024 × 512-pixel input images.
Figure 3 presents the qualitative evaluation results
for the semantic segmentation task. The qualitative
assessment indicated that conventional methods
Zero-DCE and LLFLow, which aim to improve
image quality for human vision, produced visually
appealing enhancement results but led to incorrect
recognition outcomes. IALM also exhibited instances
of erroneous recognition. While the images enhanced
by our proposed method did not achieve visually
appealing results, they demonstrated an improvement
in recognition performance.
These experiments collectively validated the
effectiveness of our proposed method and its
generalizability across different tasks.
4.5 Ablation Study
GEM performs exposure correction and color balance
adjustment by adaptively multiplying a constant
value for each channel of the input image. However,
this simple image processing can be implicitly
modeled by multiple convolution operations by the
Table 3: Quantitative evaluation results of the semantic
segmentation task.
Model mIoU ↑
DeepLabV3+ 18.4
LLFLow and
DeepLabV3+
15.3
Zero-DCE and
DeepLabV3+
16.7
IALM and
DeepLabV3+
18.7
Ours and
DeepLabV3+
34.4
Table 4: Comparison of the processing speed of each
method in the semantic segmentation task. Input images are
RGB images of size 1024 × 512-pixel in this task.
Model Params [M] Latency [ms]
DeepLabV3+ 5.2 18.1
LLFLow and
DeepLabV3+
44.1 1,027.9
Zero-DCE and
DeepLabV3+
5.3 38.2
IALM and
DeepLabV3+
5.7 32.1
Ours and
DeepLabV3+
5.8 26.3
PAM at a later stage. In that case, it may not be
necessary to include GEM before PAM. We
considered this possibility and conducted an ablation
study to validate the effectiveness of each of the two
proposed modules.
In this experiment, we adopted the semantic
segmentation task as a recognition task and
DeepLabV3+ as a downstream recognition model and
reported the mIoU when the proposed method was
trained excluding GEM and PAM, respectively.
If PAM can approximate the image processing
performed by GEM, we predicted that the mIoU,
when trained with PAM alone, would be consistent
with that when trained with both GEM and PAM.
We show an experimental result in Table 5.
According to Table 5, training only GEM, while
omitting the PAM, yields superior improvement in
the performance of the downstream recognition
model compared to conventional methods. However,
the performance improvement was still marginal,
indicating the limitation of the performance
improvement with adopting only global adjustment of
the input image.
VISAPP 2025 - 20th International Conference on Computer Vision Theory and Applications
228
(a) Comparison of enhanced images obtained by each low-light image enhancement method with the predicted results.
(b) Prediction results for images enhanced by each method.
Figure 3: Qualitative evaluation results of the semantic segmentation task.
Recognition-Oriented Low-Light Image Enhancement Based on Global and Pixelwise Optimization
229
On the other hand, when only the PAM was
trained without the GEM, the accuracy is
dramatically improved. This demonstrated the
advantage of adopting pixelwise image correction.
However, the performance when only PAM was
trained did not achieve the performance when both
GEM and PAM were trained. This demonstrated the
advantage of explicitly splitting and introducing
Global image adjustment and Pixelwise Image
Adjustment.
In principle, replacing PAM with a deeper fully
convolutional neural network could implicitly
approximate the processing of GEM in the previous
stage. However, by making global image adjustment
and pixelwise adjustment independent, a lightweight
convolutional neural network alone can effectively
improve the performance of the downstream
recognition model.
Through the above experiment, we have
demonstrated the effectiveness of the two proposed
modules.
Table 5: Quantitative results of the ablation study.
Model mIoU ↑
DeepLabv3+ 18.4
Ours (w/o PAM) and
DeepLabv3+
21.9
Ours (w/o GEM) and
DeepLabv3+
31.6
Ours and
DeepLabv3+
34.4
5 DISCUSSIONS
We proposed a novel low-light image enhancement
method aimed at improving the performance of
downstream recognition models. To evaluate the
effectiveness and generalizability of our proposed
method, we conducted experiments using two
recognition tasks: single person pose estimation and
semantic segmentation.
The experimental results indicate that the images
corrected by our proposed method may contain
artifacts, and thus may not be suitable for human
vision. However, they were shown to improve
recognition performance compared to other methods.
This suggests that high image quality may not
necessarily required to improve recognition
performance.
Furthermore, comparing the images corrected by
the proposed method for the two tasks, it was unclear
which factors mainly contributed to the improved
performance. When the recognition model by Lee et
al. for pose estimation was used, purple artifacts
appeared in the output images. In contrast, using
DeepLabV3+ as the recognition model resulted in
output images with a greenish tint. The images
corrected by our method had unique characteristics
for each downstream recognition model, but these
characteristics were not maintained when the
recognition model was changed.
These findings suggest that the images
comprehensible for recognition models may not share
a universal feature, but images corrected for each
recognition model end up having their own set of
unique features corresponding to the respective
recognition model.
6 CONCLUSIONS
We proposed a novel low-light image enhancement
method aimed at improving the performance of image
recognition models based on neural networks. Our
proposed method consists of two modules: the Global
Enhance Module (GEM) and the Pixelwise
Adjustment Module (PAM). The GEM derives
exposure correction parameters from resized low-
resolution images to adjust the exposure and the color
balance of the input image globally. The PAM takes
the exposure-corrected image from the GEM and
predicts pixel-level features that are effective for
downstream recognition models, thereby enhancing
the image effectively. The entire framework is trained
end-to-end, optimizing the low-light image
enhancement model to minimize model specific loss,
enhancing its performance.
Notably, our proposed method is composed
entirely of a lightweight convolutional neural network,
containing only 577k parameters. This makes it
significantly lighter compared to conventional low-
light image enhancement methods. As a result, using
our method as a front-end filter allows for the
improvement of recognition performance under low-
light conditions without the need to retrain various
existing pretrained recognition models. Experiments
across two tasks utilizing low-light images
demonstrated that applying our method to recognition
models achieved higher performance than
conventional methods, confirming its effectiveness
and generalizability.
VISAPP 2025 - 20th International Conference on Computer Vision Theory and Applications
230
REFERENCES
Andriluka, M., Pishchulin, L., Gehler, P., & Schiele, B.
(2014). 2d human pose estimation: New benchmark and
state of the art analysis. In Proceedings of the IEEE
Conference on computer Vision and Pattern
Recognition (pp. 3686-3693).
Cai, J., Gu, S., & Zhang, L. (2018). Learning a deep single
image contrast enhancer from multi-exposure images.
IEEE Transactions on Image Processing, 27(4), 2049-
2062.
Chen, C., Chen, Q., Xu, J., & Koltun, V. (2018). Learning
to see in the dark. In Proceedings of the IEEE
conference on computer vision and pattern recognition
(pp. 3291-3300).
Chen, L. C., Zhu, Y., Papandreou, G., Schroff, F., & Adam,
H. (2018). Encoder-decoder with atrous separable
convolution for semantic image segmentation. In
Proceedings of the European conference on computer
vision (ECCV) (pp. 801-818).
Cordts, M., Omran, M., Ramos, S., Rehfeld, T., Enzweiler,
M., Benenson, R., ... & Schiele, B. (2016). The
cityscapes dataset for semantic urban scene
understanding. In Proceedings of the IEEE Conference
on Computer Vision and Pattern Recognition, 3213-
3223.
Garcia-Garcia, A., Orts-Escolano, S., Oprea, S., Villena-
Martinez, V., Martinez-Gonzalez, P., & Garcia-
Rodriguez, J. (2018). A survey on deep learning
techniques for image and video semantic
segmentation. Applied Soft Computing, 70, 41-65.
Geiger, A., Lenz, P., & Urtasun, R. (2012, June). Are we
ready for autonomous driving? the kitti vision
benchmark suite. In 2012 IEEE conference on
computer vision and pattern recognition (pp. 3354-
3361). IEEE.
Guo, C., Li, C., Guo, J., Loy, C.C., Hou, J., Kwong, S., &
Cong, R. (2020). Zero-reference deep curve estimation
for low-light image enhancement. In Proceedings of the
IEEE/CVF Conference on Computer Vision and
Pattern Recognition, 1780-1789.
Güler, R. A., Neverova, N., & Kokkinos, I. (2018).
Densepose: Dense human pose estimation in the wild.
In Proceedings of the IEEE conference on computer
vision and pattern recognition (pp. 7297-7306).
Lee, S., Rim, J., Jeong, B., Kim, G., Woo, B., Lee, H., &
Kwak, S. (2023). Human pose estimation in extremely
low-light conditions. In Proceedings of the IEEE/CVF
Conference on Computer Vision and Pattern
Recognition, 704-714.
Li, C., Guo, C., Han, L., Jiang, J., Cheng, M. M., Gu, J., &
Loy, C. C. (2021). Low-light image and video
enhancement using deep learning: A survey. IEEE
transactions on pattern analysis and machine
intelligence, 44(12), 9396-9416.
Liang, J., Wang, J., Quan, Y., Chen, T., Liu, J., Ling, H., &
Xu, Y. (2021). Recurrent exposure generation for low-
light face detection. IEEE Transactions on
Multimedia, 24, 1609-1621.
Lin, T. Y., Maire, M., Belongie, S., Hays, J., Perona, P.,
Ramanan, D., ... & Zitnick, C. L. (2014). Microsoft
coco: Common objects in context. In Computer Vision–
ECCV 2014: 13th European Conference, Zurich,
Switzerland, September 6-12, 2014, Proceedings, Part
V 13 (pp. 740-755). Springer International Publishing.
Liu, J., Xu, D., Yang, W., Fan, M., & Huang, H. (2021).
Benchmarking low-light image enhancement and
beyond.
International Journal of Computer Vision, 129,
1153-1184.
Muhammad, K., Hussain, T., Ullah, H., Del Ser, J., Rezaei,
M., Kumar, N., ... & de Albuquerque, V. H. C. (2022).
Vision-based semantic segmentation in scene
understanding for autonomous driving: Recent
achievements, challenges, and outlooks. IEEE
Transactions on Intelligent Transportation
Systems, 23(12), 22694-22715.
Ogino, Y., Shoji, Y., Toizumi, T., & Ito, A. (2024). ERUP-
YOLO: Enhancing Object Detection Robustness for
Adverse Weather Condition by Unified Image-
Adaptive Processing. arXiv preprint arXiv:2411.02799.
Ono, S., Ogino, Y., Toizumi, T., Ito, A., & Tsukada, M.
(2024). Improving Low-Light Image Recognition
Performance Based on Image-Adaptive Learnable
Module. In The 19th International Conference on
Computer Vision Theory and Applications, vol. 3, 721-
728.
Ronneberger, O., Fischer, P., & Brox, T. (2015). U-net:
Convolutional networks for biomedical image
segmentation. In Medical Image Computing and
Computer-Assisted Intervention–MICCAI 2015: 18th
International Conference, Munich, Germany, October
5-9, 2015, Proceedings, Part III, 234-241. Springer
International Publishing.
Tan, X., Xu, K., Cao, Y., Zhang, Y., Ma, L., & Lau, R.W.
(2021). Night-time scene parsing with a large real
dataset. IEEE Transactions on Image Processing, 30,
9085-9098.
Tian, Z., Qu, P., Li, J., Sun, Y., Li, G., Liang, Z., & Zhang,
W. (2023). A survey of deep learning-based low-light
image enhancement. Sensors, 23(18), 7763.
Wang, H., Chen, Y., Cai, Y., Chen, L., Li, Y., Sotelo, M.
A., & Li, Z. (2022). SFNet-N: An improved SFNet
algorithm for semantic segmentation of low-light
autonomous driving road scenes. IEEE Transactions on
Intelligent Transportation Systems, 23(11), 21405-
21417.
Wang, W., Wu, X., Yuan, X., & Gao, Z. (2020). An
experiment-based review of low-light image
enhancement methods. Ieee Access, 8, 87884-87917.
Wang, X., Zhao, Y., & Pourpanah, F. (2020). Recent
advances in deep learning. International Journal of
Machine Learning and Cybernetics, 11, 747-750.
Wang, Y., Wan, R., Yang, W., Li, H., Chau, L.P., & Kot,
A. (2022). Low-light image enhancement with
normalizing flow. In Proceedings of the AAAI
Conference on Artificial Intelligence, vol. 36, no. 3,
2604-2612.
Recognition-Oriented Low-Light Image Enhancement Based on Global and Pixelwise Optimization
231
Wei, C., Wang, W., Yang, W., & Liu, J. (2018). Deep
retinex decomposition for low-light
enhancement. arXiv preprint arXiv:1808.04560.
Zheng, C., Wu, W., Chen, C., Yang, T., Zhu, S., Shen, J., ...
& Shah, M. (2023). Deep learning-based human pose
estimation: A survey. ACM Computing Surveys, 56(1),
1-37.
VISAPP 2025 - 20th International Conference on Computer Vision Theory and Applications
232
