FFAD: Fixed-Position Few-Shot Anomaly Detection for Wire Harness

Utilizing Vision-Language Models

Powei Liao

, Pei-Chun Chien

, Hiroki Tsukida

, Yoichi Kato

and Jun Ohya

Department of Modern Mechanical and Engineering, Waseda University, Tokyo, Japan

AI Digital Division, Yazaki Corporation, Tokyo, Japan

Global Center for Science and Engineering, Waseda University, Tokyo, Japan

Keywords:

Few-Shot Learning, Anomaly Detection, Vision-Language Model.

Abstract:

Anomaly detection in wire harness assembly for automobiles is a challenging task due to the deformable

nature of cables and the diverse assembly environments. Traditional deep learning methods require large

datasets, which are difﬁcult to obtain in manufacturing settings. To address these challenges, we propose

Fixed-Position Few-Shot Anomaly Detection (FFAD), a method that leverages pre-trained vision-language

models, speciﬁcally CLIP, to perform anomaly detection with minimal data. By capturing images from ﬁxed

positions and using position-based learnable prompts and visual augmentation, FFAD can detect anomalies in

complex wire harness situations without the need for extensive data collection. Our experiments demonstrated

that FFAD achieves over 90% accuracy with fewer than 16 shots per class, outperforming existing few-shot

learning methods.

1 INTRODUCTION

With the rapid advancement of deep learning, manu-

facturing around the world is becoming increasingly

intelligent through integrating complex sensors and

the use of Internet of Things (IoT) (Wang et al.,

2018). The manufacturing sector has also seen a

steady increase in the adoption of intelligent indus-

trial robots, where their performance and problem-

solving ability have been substantially improved year-

by-year (Benotsmane et al., 2020). One can conﬁ-

dently state that automation technologies, especially

robotics and digital solutions, are reshaping the pro-

duction landscape as we know it, making these pro-

cesses more ﬂexible and efﬁcient than ever before

(Karabegovi

c et al., 2018; Papulov

a et al., 2022).

However, despite the signiﬁcant advancement in au-

tomation technologies, the “nerves and blood vessels”

of automobiles – wire harness – is still mostly handled

and manufactured by humans due to inherent difﬁcul-

ties (Heisler et al., 2021).

Broadly speaking, a wire harness functions as

a bundle of cables that transmit power and electri-

cal signals between various parts of an automobile.

∗

The work described by this paper is a collaborative

research between Yazaki Corp. and Waseda University.

As automobiles become more sensor-reliant and with

electric vehicles (EVs) receive more attention, the de-

mand for wire harnesses of greater complexity con-

tinues to rise. However, given the declining avail-

ability of labour in many regions of the world, the

automation of wire harness manufacturing is rapidly

emerging as a prominent area of research (Navas-

Reascos et al., 2022). One particular interest is in the

ability to perform anomaly detection and quality as-

surance with deep-learning models. Such a system

would require a vision system capable of detecting

various anomalies during the manufacturing process.

With effective anomaly detection, the system can in-

tervene to halt an operation when necessary, thereby

preventing damage such as wire entanglement or for-

eign object intrusion. This also opens the possibility

for auto rectiﬁcation to be performed.

Currently, anomaly detection for wire harnesses

remains challenging for several reasons. Firstly, ca-

bles – also referred to as deformable linear objects

(DLOs) – are non-rigid and highly deformable, mean-

ing their shape can change drastically during the

manufacturing process, complicating detection efforts

(Zhou et al., 2020). Secondly, the deﬁnition of an

anomaly can often be ambiguous as only parts of wire

harness are anomalous while the overall appearance

remains satisfactory. Thirdly, the layout of a wire har-

Liao, P., Chien, P.-C., Tsukida, H., Kato, Y. and Ohya, J.

FFAD: Fixed-Position Few-Shot Anomaly Detection for Wire Harness Utilizing Vision-Language Models.

DOI: 10.5220/0013164400003905

Paper published under CC license (CC BY-NC-ND 4.0)

In Proceedings of the 14th International Conference on Pattern Recognition Applications and Methods (ICPRAM 2025), pages 647-656

ISBN: 978-989-758-730-6; ISSN: 2184-4313

647

Figure 1: Overview of our application scenario. The model will not be trained throughout the entire process; instead, a pair

of prompts will be trained at each position to detect anomalies.

ness can vary signiﬁcantly for each scenario, leading

to diverse and unpredictable assembly environments.

Finally, as with other industrial anomaly detection

tasks, most of the deep learning methods heavily de-

pend on available datasets. The high demand for both

the quantity and quality of datasets makes it challeng-

ing to reproduce the same results on varying produc-

tion lines.

To address these challenges, we propose a

method – Fixed-Position Few-Shot Anomaly Detec-

tion (FFAD) – which aims to detect user-deﬁned

anomalies in complex wire harness situations without

requiring a large amount of data for learning and can

be easily adapted to different environments by lever-

aging the power of vision-language models.

Few-shot learning, which enables a model to per-

form downstream tasks with only a small amount of

data (typically fewer than 16 shots), has been gain-

ing increasing attention (Wang et al., 2020). CLIP

(Contrastive Language-Image Pre-Training) (Radford

et al., 2021), a vision-language model pre-trained on

400 million image-text pairs, aligns images and texts

in a shared embedding space, enabling it to associate

visual and textual concepts. This architecture allows

CLIP to demonstrate remarkable zero-shot learning

capabilities. The extensive pre-trained knowledge of

CLIP can be easily adapted to downstream tasks; for

instance, CoOp (Zhou et al., 2022b) adapts CLIP for

downstream classiﬁcation tasks by making the text in-

put prompt learnable, then performs training on a few

samples from the target tasks. Building on a simi-

lar concept as CoOp, we propose our method FFAD

that leverages CLIP’s pre-trained knowledge to per-

form few-shot learning for anomaly detection in wire

harnesses.

Figure 1 shows an overview of our applica-

tion scenario. In traditional deep learning meth-

ods, the entire pre-trained model is ﬁne-tuned with a

large amount of additional data to detect all possible

anomalies, which requires substantial labor and time.

Furthermore, if an additional anomaly class is to be

added, the whole model needs to be ﬁne-tuned again.

To address this, the FFAD approach trains a speciﬁc

prompt for each position on the assembly board, al-

lowing for detection of different anomalies without

the need of full model ﬁne-tuning. Additionally, we

introduce Position-based Visual Augmentation to fur-

ther enhance the model’s anomaly detection capabili-

ties. Our experiments show that FFAD outperformed

other few-shot learning methods, achieving over 90%

accuracy with 16 shots per class, less than 2 minutes

of training time, and requiring less than 200 KB to

store each prompt.

The main contributions of this paper are as fol-

lows:

• Proposed a novel wire harness anomaly detection

method – FFAD, and introduced Position-Based

Visual Augmentation to enhance model accuracy.

• Development of a wire harness anomaly dataset

for detection tasks and demonstrated FFAD’s su-

periority over existing methods through extensive

experiments.

• Ablation studies were carried out to investigate

the impact of visual prompt engineering and

model scaling on FFAD’s performance.

ICPRAM 2025 - 14th International Conference on Pattern Recognition Applications and Methods

648

2 RELATED WORK

2.1 Anomaly Detection for Wire

Harness

Wire harness is an essential component of automo-

biles, connecting various parts throughout the vehicle.

It transmits power and information between multiple

components, ranging from assistive systems to safety-

critical systems. Due to its importance, up to 90% of

the wire harness manufacturing process still requires

human involvement (Nguyen et al., 2021). There-

fore, automating wire harness production while main-

taining quality and reliability has become a signiﬁ-

cant focus in recent years (Navas-Reascos et al., 2022;

Trommnau et al., 2019; Hermansson et al., 2013).

For such process to be automated, a vision system

capable of understanding the assembly scene and de-

tecting anomalies within said scene is crucial (Wang

et al., 2024). Li et al. (2024) developed a mobile-

based visual imaging system using YOLOv5 for real-

time detection and recognition, which is used to error-

proof the installation process of automotive wiring

harness relays. Nguyen et al. (2022) proposed a deep

learning-based data processing pipeline that utilizes

both real and synthetic point clouds for automated in-

spection. Chien et al. (2024) introduced a method for

classifying different cable tendencies, enabling fur-

ther anomaly detection through semantic segmenta-

tion by utilizing real and simulated RGB data.

While these methods can detect anomalies in the

wire harness process using deep learning techniques,

they often rely on large datasets and may require syn-

thetic data to achieve optimal performance. Our pro-

posed method requires only a small amount of data

by leveraging the pre-trained knowledge of vision-

language models.

2.2 Few-Shot Learning Utilizing

Vision-Language Models

Following the remarkable success of CLIP (Radford

et al., 2021), there has been a rising interest in re-

search on vision-language models. Numerous down-

stream tasks based on CLIP have emerged in recent

years, including multi-label classiﬁcation (Sun et al.,

2022; Liao and Nakano, 2025), object detection (Gu

et al., 2021; Wang et al., 2022), image segmenta-

tion (Liang et al., 2023; L

uddecke and Ecker, 2022;

Xie et al., 2022), image editing (Kwon and Ye, 2022;

Wei et al., 2022), and image captioning (Hessel et al.,

2021; Mokady et al., 2021).

Among the various studies, signiﬁcant research

has been devoted to utilizing CLIP for few-shot learn-

Figure 2: RealSense D405 camera mounted on UR5e robot.

Table 1: The number of images from different position.

Position Normal Anomaly

Connector 80 51

Cable Holder 80 51

Cable 1 65 66

Cable 2 63 65

ing. CoOp (Zhou et al., 2022b) treats the prompt in

the input text as a set of learnable vectors, enabling

few-shot learning by directly training the prompt. Co-

CoOp (Zhou et al., 2022a), on the other hand, takes

a different approach by training a lightweight net-

work that uses the input image to adjust the prompt

adaptively, rather than training it directly. CLIP-

Adapter (Gao et al., 2024) enhances the pre-trained

CLIP image and text encoders by adding three fully

connected layers, performing few-shot learning by

training only the additional layers. Shtedritski et al.

(2023) demonstrates that visual prompt engineering,

such as simple image edits like drawing a red cir-

cle, can improve CLIP’s performance on complex

tasks. Building on the concepts of CoOp and visual

prompt engineering, our proposed method adapts the

pre-trained CLIP model for wire harness assembly

anomaly detection using few-shot learning. Addition-

ally, we incorporate Position-Based Visual Augmen-

tation to further enhance the model’s performance.

3 DATASET

We constructed a wire harness assembly dataset to

demonstrate the effectiveness of our method. As

shown in Figure 2, we mounted a camera (Intel Re-

alSense D405) onto a robot arm (Universal Robots

UR5e) to capture ﬁxed-position images. For each

capture, the wire harness is automatically assembled

using the robot arm by traversing through registered

FFAD: Fixed-Position Few-Shot Anomaly Detection for Wire Harness Utilizing Vision-Language Models

649

Normal Anomaly Normal(Augmented) Anomaly(Augmented)

ConnectorCable 1Cable HolderCable 2

Figure 3: Examples of each position from the dataset before and after visual augmentation. (The cable holder is blurred for

conﬁdentiality reasons).

coordinates. Speciﬁcally, the robot arm ﬁrst grabs the

connector from the starting base, attaches the cables

to the cable holder – a partially enclosed C-shaped ca-

ble supporter, and places the connector onto the goal

base. Then, images of the assembled wire harness are

captured from different ﬁxed positions. As illustrated

in Figure 1, different positions have different anoma-

lies that need to be detected. For wire harnesses, some

common anomalies to be identiﬁed include, but are

not limited to:

• if the connector is appropriately ﬁxed to its base

• if the cable is routed through the cable holder

• if the cable is situated within a certain region

In this dataset, images are taken from four different

positions: the connector, the cable holder, and two

positions along the cables. Table 1 shows the num-

ber of normal and anomalous images for each posi-

tion. Figure 3 provides examples of both normal and

anomalous situations for each position.

4 METHODOLOGY

The structure of the proposed Fixed-position Few-

shot Anomaly Detection (FFAD) is depicted in Fig-

ure 4. In our setup, the camera is mounted on the

robot arm and moves in conjunction with the robot’s

ICPRAM 2025 - 14th International Conference on Pattern Recognition Applications and Methods

650

Figure 4: Illustration of our proposed approach FFAD.

movements. This arrangement ensures that the ”ﬁxed-

position” requirement of FFAD can be effectively met

by registering robot coordinates. By capturing images

at the same coordinates, the camera’s position and an-

gle are guaranteed to be identical. FFAD takes two

inputs: the original image and a Position ID, which

identiﬁes the location where the image was captured.

Based on the Position ID, we create two lists,

Position-Based Prompts List and Position-Based Fea-

ture Points List. In Position-Based Feature Points

List, four feature points are stored for each position.

These feature points are manually crafted for use in

the Position-Based Visual Augmentation.

The visual augmentation process uses OpenCV to

draw a quadrilateral on the original image based on

these feature points, allowing the model to better fo-

cus on the relevant area of the quadrilaterals. As

shown in Figure 3, each position has a speciﬁc vi-

sual augmentation policy. As for the connector, the

goal is to detect whether it is properly set on the base.

Thus, the feature points of quadrilateral are instead

deﬁned along the edges of the connector and the base.

In the anomaly example of connector shown in Fig-

ure 3, distinguishing between normal and anomalous

situations is difﬁcult when the visual augmentation

has not been applied. However, after applying vi-

sual augmentation, it is noticeable that the connector

is slightly misaligned to the left of the base.

Similarly for the cable position, we deﬁne a

quadrilateral around the cables as the appropriate re-

gion of which cables are to exist within. Any loose

cable that falls out of this region is to be classiﬁed as

lack of tension. Lastly, for the cable holder position,

a quadrilateral is drawn around the holder to highlight

the key area for detecting whether the cables are prop-

erly routed through the holder.

In Position-Based Prompts List, each position is

associated with a pair of learnable prompts: prompt

for normal, Prompt

, and prompt for anomaly,

Prompt

. There are N learnable vectors V

,· · · ,V

each prompt with the word “Normal” and “Anomaly”,

these prompts can be written as:

Prompt

= [V

id N

,...,V

id N

, Normal ] (1)

Prompt

= [V

id A

,...,V

id A

, Anomaly] (2)

These vectors are input into the pre-trained CLIP text

encoder E

, producing two Text-Feature Vectors T

and T

= E

(Prompt

) (3)

= E

(Prompt

) (4)

On the other side, the Position-Based Augmented im-

age is input into the pre-trained CLIP image encoder

, yielding the Image Feature Vector I:

I = E

(Augmented Image) (5)

Subsequently, the Image Feature Vector I is compared

to both Text-Feature Vectors, T

and T

, using cosine

similarity S

cos

to produce normal and anomaly logits:

= S

cos

(I,T

), S

= S

cos

(I,T

) (6)

Finally, with the pair of logits, the binary classiﬁca-

tion output p is given by:

p =

exp(S

/τ)

exp(S

/τ) + exp(S

/τ)

(7)

FFAD: Fixed-Position Few-Shot Anomaly Detection for Wire Harness Utilizing Vision-Language Models

651

(a) (b) (c)

(d) (e)

Figure 5: Results of few-shot learning of the deﬁned four positions.

5 EXPERIMENTS

5.1 Few-Shot Anomaly Detection for

Wire Harness

5.1.1 Training Setting

To evaluate the effectiveness of our method for

anomaly detection in wire harnesses, we trained and

tested it on the dataset described in Section 3. We

trained FFAD under the few-shot setups of 1, 2, 4, 8,

16 shots. These training examples were randomly se-

lected from the dataset and served as training and val-

idation sets, and the remaining data was used for test-

ing. In the Position-Based Learnable Prompts, we use

8 learnable vectors for each prompt, sufﬁxed with the

class name “Normal” or “Anomaly” for each position.

We employed ViT-B/16 (Dosovitskiy, 2020) as our

choice of the pre-trained CLIP image encoder, as well

as the use of pre-trained models provided by Ope-

nAI.

The entire pre-trained model is kept untrained,

while only the learnable prompts are updated during

training. We did not use data augmentation technique

but used only normalization to preserve ﬁxed-position

information as much as possible. We used cross-

entropy as the loss function and SGD as the optimizer.

The model was trained for 200 epochs with a learning

https://github.com/openai/CLIP

rate of 0.002. The overall implementation is based on

CoOp’s open-source code.

5.1.2 Baselines

For our task, every single position can be regarded as

a binary classiﬁcation task. We compare our method

against one zero-shot method, CLIP, and three few-

shot methods: CoOp, CoCoOp, and CLIP-Adapter.

Both CoOp and CoCoOp utilize 8 learnable vectors,

which is consistent with our approach. All few-shot

methods follow the same training setup as FFAD and

utilize the ViT-B/16 pre-trained model from OpenAI.

5.1.3 Experiments Result & Analysis

The plot in Figure 5 shows the accuracy for four dif-

ferent positions as well as the average accuracy across

these positions. For few-shot learning methods, the

result is the average over three separate runs. As de-

picted in Figure 5(a), FFAD outperforms the other

three methods when averaged over 4 positions in ev-

ery few-shot setup.

Figure 5(b) shows that FFAD outperforms the

other methods at the “Connector” position when the

few-shot setup is of 8 or 16 shots. The visual differ-

ence between normal and anomaly is relatively subtle

when no adequate visual augmentation is applied, as

seen in Figure 3, which may be challenging for the

https://github.com/KaiyangZhou/CoOp

ICPRAM 2025 - 14th International Conference on Pattern Recognition Applications and Methods

652

Table 2: Result of different visual prompt engineering.

Color Thickness Connector Cable 1 Cable Holder Cable 2 Average

Red 1 81.70 92.92 75.65 89.84 85.03

Red 2 82.85 90.69 76.30 90.27 85.00

Green 1 81.39 94.71 78.93 87.42 85.61

Blue 1 79.92 91.54 77.13 90.16 84.69

Color = ‘Red’, Thickness = 2 Color = ‘Green’, Thickness = 1 Color = ‘Blue’, Thickness = 1

(a) (b) (c)

Figure 6: Examples of visual prompt engineering with different settings.

model to distinguish differences. CoCoOp performs

well with fewer shots but is unstable and does not per-

form as effectively when the few-shot setups increase

to 8 or 16 shots.

Figure 5(c) plot demonstrates that FFAD signiﬁ-

cantly outperforms the other methods at the “Cable

1” position, highlighting the effectiveness of visual

augmentation.

In Figure 5(d), FFAD shows a slight advantage

over the other methods at the “Cable Holder” po-

sition, indicating that visual augmentation aids the

model focusing on the relationship between the cable

and the cable holder, even though it is not as intuitive

as in other positions.

However, Figure 5(e) presents unexpected results

where FFAD does not surpass the other methods at

the “Cable 2” position. This is likely due to the rel-

ative simplicity of this task, as methods without vi-

sual augmentation were already achieving 100% ac-

curacy with 8 and 16 shots. As shown in Figure 3, the

background of “Cable 2” is relatively clean, the ap-

plication of visual augmentation may have introduced

noise, leading to a negative effect in setups with fewer

shots.

Zero-shot CLIP did not perform well across all

four positions, likely due to the speciﬁc examples in

our task not being available in the web-scraped data

that CLIP was originally trained on. Additionally,

using the terms ‘Normal’ and ‘Anomaly’ to describe

the two classes may have posed a challenge for the

model to comprehend when no additional training is

done. Meanwhile, all few-shot methods demonstrated

that with proper instruction, the pre-trained knowl-

edge can be effectively adapted to the downstream

task. This suggests that a small amount of task-

speciﬁc guidance can signiﬁcantly enhance model

performance, leveraging the underlying capabilities

of the pre-trained model.

Overall, the results from Figure 5 show that inte-

grating visual augmentation allows FFAD to improve

anomaly detection performance across different posi-

tions and few-shot setups.

5.2 Ablation Study

To further assess the impact of the FFAD components,

an ablation study is carried out to examine the roles of

”Visual Prompt Engineering” and ”Model Scaling”,

aiming to gain deeper insights into their inﬂuence on

the performance of our few-shot anomaly detection

method.

5.2.1 Visual Prompt Engineering

Inspired by visual prompt engineering (Shtedritski

et al., 2023), we propose Position-Based Visual Aug-

mentation in FFAD, where a red quadrilateral is

drawn on the original image to emphasize region of

interest thus enhance the performance. To investigate

how different colors or thicknesses of the quadrilat-

eral in visual augmentation affect detection results,

we conducted experiments with various settings. Ta-

ble 2 presents the results of our model under differ-

ent visual augmentation conﬁgurations, averaged over

few-shot settings of 1, 2, 4, 8, and 16 shots, with 3

runs for each setting. Figure 6 shows examples of vi-

sual prompt engineering with different settings.

By adjusting the thickness to 2 pixels while keep-

FFAD: Fixed-Position Few-Shot Anomaly Detection for Wire Harness Utilizing Vision-Language Models

653

Table 3: Average result of few-shot learning on different

model scaling.

Models Average Accuracy

ViT-B/32@224px 85.32

ViT-B/16@224px 85.03

ViT-L/14@224px 87.70

ViT-L/14@336px 86.05

Figure 7: Result of few-shot learning on different model

scaling.

ing the color as “red”, there is no signiﬁcant change in

performance. However, a slight decrease in accuracy

at position “Cable 2” was observed. This may be due

to the thicker red quadrilateral resembling the yellow

tape in the background, and the thicker line may occa-

sionally occlude the cable, as is shown in Figure 6(a),

causing confusion for the model.

Changing the color to “green” while keeping the

thickness at 1 pixel, we observed an increase in per-

formance at the “Cable 1” and “Cable Holder” posi-

tions but again a decrease at the “Cable 2” position.

This may be caused by the green quadrilateral having

more contrast with the backgrounds of “Cable 1” and

“Cable Holder” compared to the red quadrilateral, as

is shown in Figure 6(b), while there is less contrast at

“Cable 2” position.

A similar effect was observed with the color blue,

where it blended into the metallic texture of the con-

nector, as is shown in Figure 6(c), resulting in a per-

formance drop at the “Connector” position. However,

due to its higher contrast with the cable holder, the

performance improved at the “Cable Holder” posi-

tion.

In summary, the color of the quadrilateral and its

contrast with the background may have some impact

on performance. Choosing suitable colors for dif-

ferent positions that improve contrast could help en-

hance model performance.

5.2.2 Model Scaling

Another important factor to performance is the model

scaling of the pre-trained CLIP model. To understand

how different model sizes impact our task, four differ-

ent sizes of Vision Transformer (ViT) obtained from

OpenAI’s ofﬁcial CLIP implementation were evalu-

ated. Table 3 presents the results for these models.

Each result is averaged across four different positions,

where each position is averaged over ﬁve different

few-shot setups: 1, 2, 4, 8, and 16 shots. Each few-

shot setup result is again averaged over three runs. Ta-

ble 3 indicates that increasing the size of the vision

transformer model from B (Base) to L (Large) results

in a slight improvement in accuracy. However, chang-

ing the patch size of the ViT-B model from 32 to 16

did not signiﬁcantly affect performance.

Interestingly, for the ViT-L/14 model, increasing

the input image resolution from 224×224 to 336×336

unexpectedly decreases the performance. This could

be due to that, in our task, the region of interest for

anomaly detection has been zoomed in. As a result,

the anomalies already occupy a signiﬁcant portion of

the input image, and a resolution of 224×224 is suf-

ﬁcient to capture the necessary details. Increasing

the input resolution further might introduce unnec-

essary complexity, making the model harder to train

and demanding more computational resources with-

out yielding much improvement.

Figure 7 shows the accuracy plot under differ-

ent few-shot conﬁgurations. The ViT model with a

Large size outperforms the Base size when trained

with fewer shots, such as 1 and 4. However, the

performance difference diminishes as the number of

shots increases to 8 or 16, where both models achieve

nearly the same accuracy.

6 CONCLUSIONS

In this paper, we proposed FFAD, a Fixed-Position

Few-Shot Anomaly Detection method for wire har-

ness assembly, which leverages pre-trained vision-

language models, speciﬁcally CLIP, to perform

anomaly detection with minimal data. By capturing

images from ﬁxed positions and employing position-

based learnable prompts along with visual augmenta-

tion, FFAD can effectively detect anomalies in com-

plex wire harness situations without the need for

extensive data collection. Our experimental results

demonstrated that FFAD outperforms other few-shot

learning methods, and achieves over 90% accuracy

with fewer than 16 shots per class. The ablation stud-

ies indicated that visual prompt engineering, such as

ICPRAM 2025 - 14th International Conference on Pattern Recognition Applications and Methods

654

adjusting the color and thickness of the augmentation,

can impact the model’s performance. Additionally,

model scaling shows that larger pre-trained models

can improve accuracy, especially in low-shot scenar-

ios. We hope that the ideas and empirical ﬁndings

presented in this paper provide valuable insights for

unlocking potential industrial applications that only

have access to limited data. In the future, we plan to

enable the model to learn the visual augmentation pol-

icy from the data itself, rather than relying on manual

decisions, in order to simplify the overall process.

REFERENCES

Benotsmane, R., Dud

as, L., and Kov

acs, G. (2020). Survey

on new trends of robotic tools in the automotive in-

dustry. In Vehicle and automotive engineering, pages

443–457. Springer.

Chien, P.-C., Liao, P., Fukuzawa, E., and Ohya, J. (2024).

Classifying cable tendency with semantic segmenta-

tion by utilizing real and simulated rgb data. In Pro-

ceedings of the IEEE/CVF Winter Conference on Ap-

plications of Computer Vision, pages 8430–8438.

Dosovitskiy, A. (2020). An image is worth 16x16 words:

Transformers for image recognition at scale. arXiv

preprint arXiv:2010.11929.

Gao, P., Geng, S., Zhang, R., Ma, T., Fang, R., Zhang,

Y., Li, H., and Qiao, Y. (2024). Clip-adapter: Bet-

ter vision-language models with feature adapters. In-

ternational Journal of Computer Vision, 132(2):581–

595.

Gu, X., Lin, T.-Y., Kuo, W., and Cui, Y. (2021).

Open-vocabulary object detection via vision and

language knowledge distillation. arXiv preprint

arXiv:2104.13921.

Heisler, P., Utsch, D., Kuhn, M., and Franke, J. (2021).

Optimization of wire harness assembly using human–

robot-collaboration. Procedia CIRP, 97:260–265.

Hermansson, T., Bohlin, R., Carlson, J. S., and S

oderberg,

R. (2013). Automatic assembly path planning for

wiring harness installations. Journal of manufactur-

ing systems, 32(3):417–422.

Hessel, J., Holtzman, A., Forbes, M., Bras, R. L., and

Choi, Y. (2021). Clipscore: A reference-free eval-

uation metric for image captioning. arXiv preprint

arXiv:2104.08718.

Karabegovi

c, I., Karabegovi

c, E., Mahmi

c, M., and Husak,

E. (2018). Innovative automation of production pro-

cesses in the automotive industry. International Jour-

nal of Engineering.

Kwon, G. and Ye, J. C. (2022). Clipstyler: Image style

transfer with a single text condition. In Proceedings

of the IEEE/CVF Conference on Computer Vision and

Pattern Recognition, pages 18062–18071.

Li, S., Yuan, M., Wang, W., Cao, F., Shi, H., Zhang, Y.,

and Meng, X. (2024). Enhanced yolo-and wearable-

based inspection system for automotive wire harness

assembly. Applied Sciences, 14(7):2942.

Liang, F., Wu, B., Dai, X., Li, K., Zhao, Y., Zhang,

H., Zhang, P., Vajda, P., and Marculescu, D. (2023).

Open-vocabulary semantic segmentation with mask-

adapted clip. In Proceedings of the IEEE/CVF Con-

ference on Computer Vision and Pattern Recognition,

pages 7061–7070.

Liao, P. and Nakano, G. (2025). Bridgeclip: Automatic

bridge inspection by utilizing vision-language model.

In International Conference on Pattern Recognition,

pages 61–76. Springer.

uddecke, T. and Ecker, A. (2022). Image segmentation

using text and image prompts. In Proceedings of the

IEEE/CVF Conference on Computer Vision and Pat-

tern Recognition, pages 7086–7096.

Mokady, R., Hertz, A., and Bermano, A. H. (2021). Clip-

cap: Clip preﬁx for image captioning. arXiv preprint

arXiv:2111.09734.

Navas-Reascos, G. E., Romero, D., Stahre, J., and

Caballero-Ruiz, A. (2022). Wire harness assembly

process supported by collaborative robots: Literature

review and call for r&d. Robotics, 11(3):65.

Nguyen, H. G., Habiboglu, R., and Franke, J. (2022). En-

abling deep learning using synthetic data: A case

study for the automotive wiring harness manufactur-

ing. Procedia CIRP, 107:1263–1268.

Nguyen, H. G., Kuhn, M., and Franke, J. (2021). Manu-

facturing automation for automotive wiring harnesses.

Procedia Cirp, 97:379–384.

Papulov

a, Z., Ga

zov

a, A., and

Suﬂiarsk

y, L. (2022). Im-

plementation of automation technologies of industry

4.0 in automotive manufacturing companies. Proce-

dia Computer Science, 200:1488–1497.

Radford, A., Kim, J. W., Hallacy, C., Ramesh, A., Goh, G.,

Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark,

J., et al. (2021). Learning transferable visual models

from natural language supervision. In International

conference on machine learning, pages 8748–8763.

PMLR.

Shtedritski, A., Rupprecht, C., and Vedaldi, A. (2023).

What does clip know about a red circle? visual

prompt engineering for vlms. In Proceedings of the

IEEE/CVF International Conference on Computer Vi-

sion, pages 11987–11997.

Sun, X., Hu, P., and Saenko, K. (2022). Dualcoop: Fast

adaptation to multi-label recognition with limited an-

notations. Advances in Neural Information Processing

Systems, 35:30569–30582.

Trommnau, J., K

uhnle, J., Siegert, J., Inderka, R., and

Bauernhansl, T. (2019). Overview of the state of the

art in the production process of automotive wire har-

nesses, current research and future trends. Procedia

CIRP, 81:387–392.

Wang, H., Salunkhe, O., Quadrini, W., L

amkull, D., Ore,

F., Despeisse, M., Fumagalli, L., Stahre, J., and Jo-

hansson, B. (2024). A systematic literature review of

computer vision applications in robotized wire har-

ness assembly. Advanced Engineering Informatics,

62:102596.

FFAD: Fixed-Position Few-Shot Anomaly Detection for Wire Harness Utilizing Vision-Language Models

655

Wang, J., Ma, Y., Zhang, L., Gao, R. X., and Wu, D. (2018).

Deep learning for smart manufacturing: Methods

and applications. Journal of manufacturing systems,

48:144–156.

Wang, Y., Yao, Q., Kwok, J. T., and Ni, L. M. (2020). Gen-

eralizing from a few examples: A survey on few-shot

learning. ACM computing surveys (csur), 53(3):1–34.

Wang, Z., Codella, N., Chen, Y.-C., Zhou, L., Yang, J.,

Dai, X., Xiao, B., You, H., Chang, S.-F., and Yuan, L.

(2022). Clip-td: Clip targeted distillation for vision-

language tasks. arXiv preprint arXiv:2201.05729.

Wei, T., Chen, D., Zhou, W., Liao, J., Tan, Z., Yuan, L.,

Zhang, W., and Yu, N. (2022). Hairclip: Design your

hair by text and reference image. In Proceedings of

the IEEE/CVF Conference on Computer Vision and

Pattern Recognition, pages 18072–18081.

Xie, J., Hou, X., Ye, K., and Shen, L. (2022). Clims:

Cross language image matching for weakly super-

vised semantic segmentation. In Proceedings of the

IEEE/CVF Conference on Computer Vision and Pat-

tern Recognition, pages 4483–4492.

Zhou, H., Li, S., Lu, Q., and Qian, J. (2020). A practical

solution to deformable linear object manipulation: A

case study on cable harness connection. In 2020 5th

International Conference on Advanced Robotics and

Mechatronics (ICARM), pages 329–333. IEEE.

Zhou, K., Yang, J., Loy, C. C., and Liu, Z. (2022a). Con-

ditional prompt learning for vision-language models.

In Proceedings of the IEEE/CVF conference on com-

puter vision and pattern recognition, pages 16816–

16825.

Zhou, K., Yang, J., Loy, C. C., and Liu, Z. (2022b). Learn-

ing to prompt for vision-language models. Inter-

national Journal of Computer Vision, 130(9):2337–

2348.

ICPRAM 2025 - 14th International Conference on Pattern Recognition Applications and Methods

656