กลับไปที่นวัตกรรมทั้งหมด

การตรวจจับและรู้จำความเสียหายของรถยนต์ โดยใช้การเรียนรู้เชิงลึก

Car damage detection and recognition using deep learning

@คณะเทคโนโลยีสารสนเทศ

#KLLC 2024
#Industry 4.0
การตรวจจับและรู้จำความเสียหายของรถยนต์ โดยใช้การเรียนรู้เชิงลึก

รายละเอียด

This senior project aimed to develop an automated system for car damage detection, segmentation, and captioning. The project evaluated the performance of the YOLOv8 model for damage detection and explored pseudo-labeling methods using deep spectral methods to generate pseudo-labeling and find approach effectiveness pseudo-label for image segmentation to overcome limited labeled data. The results showed promising outcomes, with improvements in mean IoU for certain damage classes. The project also developed language models for damage captioning, generating relevant descriptions for damaged cars. Although there is room for further improvement, the project provided insights into model performance and demonstrated the potential for cost-effective pseudo-labeling and enhanced damage evaluation processes in the car insurance sector.

วัตถุประสงค์

Currently, the automobile insurance industry is a business that is growing in response to the
demand for car usage. In the automobile insurance sector, there are services that need to be performed
both before and after policy issuance. Examples include assessing the value of damaged car parts for
policyholders who have experienced damage and evaluating the condition of a car before it is insured.
From the given example of services provided by the automobile insurance business, it is evident that
these issues are complex and require knowledge and experience. Therefore, it is essential to have
experts involved in damage assessment or evaluating the claim value of damaged parts.
Accurately describing car damages is crucial in the automotive industry as it helps assess re-
pairs and calculate insurance claims. Currently, this task is carried out by trained experts who have
undergone specialized training. However, it can be time-consuming and prone to errors
With advancements in technology, many industries are seeking to leverage deep learning mod-
els to create business opportunities. However, one limitation is the requirement for large labeled
datasets, which can be time-consuming and labor-intensive to create. This high cost in terms of time
and resources makes it challenging to apply deep learning models to address business problems due
to data limitations.
The researchers are interested in developing an automated system that utilizes deep learning tech-
niques to detect and recognize car damage in images. This system aims to support the automobile
insurance industry in evaluating car conditions prior to insurance coverage and assessing damages to
car parts for claim processing. The objective is to streamline the assessment process, minimize er-
rors, and enhance the accuracy of damage evaluation. Additionally, the researchers intend to utilize
language-based models to describe the types of damages and the affected parts, providing explana-
tions. They also aim to devise an effective approach for generating pseudo-labels as a solution to the
limitations posed by insufficient labeled data.

ผู้จัดทำ

นนทพัทธ์ เทศปัญ
NONTHAPAHT TASPAN

#นักศึกษา

สมาชิก
บุคอรีย์ หมาดทิ้ง
BUKORREE MADTHING

#นักศึกษา

สมาชิก
กิติ์สุชาต พสุภา
Kitsuchart Pasupa

#อาจารย์

อาจารย์ที่ปรึกษา

โหวตนวัตกรรมนี้

กำลังดาวน์โหลด