International Research Journal of Engineering and Technology (IRJET)
e-ISSN: 2395-0056
Volume: 09 Issue: 06 | June 2022
p-ISSN: 2395-0072
www.irjet.net
CAR DAMAGE DETECTION USING DEEP LEARNING Dindayal Bhadrecha1, Divyesh Tharakan2, Chandrababu Godasu3, Hrushikesh Jadhav4 Department of Computer Science and Engineering, School of Engineering, MIT Art, Design and Technology University, Pune, Maharashtra - 412201 ---------------------------------------------------------------------***--------------------------------------------------------------------learning. Experiments have shown that our strategy is a Abstract - Image-based vehicle insurance processing is a key viable alternative to the one that has been proposed.
industry with a lot of possibilities for automation. We look at the subject of car damage categorization in this paper, where certain damages are classed as minor and others as serious. In several of the categories, fine-granularity is feasible. We go into the bowels of knowledge. We'll use based ways for this. We attempted to train a CNN straight at first. However, it does not work well with data due to the small number of tagged samples. The impact of pre-training in the domain is next examined, followed by fine-tuning. Finally, we put ensemble and transfer learning to the test. According to research, transfer learning outperforms domain-specific fine-tuning. We have a high level of commitment.
2. LITERATURE SURVEY Several models have been implemented for car damage detection. So, when it comes to object detection the Deep Learning [DL] has always been effective and shown promising results. One of the most accepted detection algorithms is the CNN (Convolutional Neural Network), as it executes well for computer vision tasks such as visual objects detection and recognition. Deep learning has been exceptional in image classification, with computing resources based on transfer learning solutions and extensive use of data. Pre-trained CNN models are very complex to understand. Large amounts of labelled data and computer resources are needed for supervised methods. Whereas, unsupervised pre-training techniques such as Autoencoders, proved to enhance the generalization performance of the classifier in case of a small quantity of labelled sample. For images, Convolutional Autoencoders (CAE) have been effective.
1. INTRODUCTION A lot of money is lost today in the car insurance market owing to claims leakage. The discrepancy between the actual claim and the underwriting claim is referred to as claims leakage or underwriting leakage. The amount that was paid and the amount that should have been paid if all of the industry's best practices were used. Visual examination, Validation and other techniques have been employed to reduce such impacts. They do, however, cause delays in the processing of claims. There have been initiatives by a small number of start-ups to reduce claim processing times. For automotive insurance claims, an automated method is available. An hour of processing is required. We use Convolutional Neural Networks in this paper. (CNN)-based systems for classifying various types of vehicle damage. Damage types include bumper dent, door dent, glass shatter, headlight shattered, tail lamp damaged, scratch, and smash. To our knowledge, there is no publicly available dataset on car damage classification on the other hand. As a result, we created our own set of data. The categorization task is challenging when using images from the internet and manually tagging them based on these qualities. There is a great deal of inter-class similarity, but the consequences are minor. We used to experiment with a variety of methodologies, such as direct training a CNN, pre-training a CNN with an auto-encoder, and then fine-tuning with transfer learning from large CNNs that had been taught, establishing an ensemble classifier on top of ImageNet the collection of trained classifiers. The most effective method appears to be transfer learning combined with ensemble
© 2022, IRJET
|
Impact Factor value: 7.529
One of the well-known techniques which has shown results for small labelled data is Transfer Learning. There are a bunch of CNN models trained on ImageNet. These models are available publicly such as VGG-19, VGG-16, Alex net, Inception, Resnet. Transferable feature representation learned by CNN minimizes the effect of over-fitting in case of small labelled sets. To the best of our knowledge, deep learning-based techniques have not been employed for automated car damage classification, especially for the fine granular classification.
3. DATASET DESCRIPTION Because there was no standard dataset for this topic, we decided to create our own by gathering photographs of damaged cars from the internet. After gathering a large number of them, we discovered that the dataset had numerous photographs of entirely scrambled and damaged automobiles, which had to be avoided for better model training. The second thing we saw was that the number of photographs was insufficient, thus data
|
ISO 9001:2008 Certified Journal
|
Page 2674