Vehicle Interior Detection (3)

▤ 목차

Data

728x90

Data plays a pivotal role in the success of AI-driven solutions, particularly in applications like Socar's vehicle contamination detection project. The quality, quantity, and diversity of data directly influence the performance, accuracy, and reliability of the AI model. In this project, data serves as the foundation for training, validating, and refining the deep learning model, enabling it to make informed decisions about the cleanliness and operational state of vehicles.

 

The data strategy involves not only collecting relevant images of vehicle interiors and dashboards but also ensuring the data represents real-world scenarios. This includes clean and dirty interiors, functional and faulty dashboards, as well as variations in lighting, angles, and vehicle types. Additionally, preprocessing and augmenting the data are critical steps to improve the robustness of the model. By addressing potential biases, enhancing diversity, and simulating various conditions, the data ensures the AI system can generalize effectively and maintain consistent performance when deployed at scale. This section will delve into the processes of data collection and augmentation, outlining their importance and methodology in the context of the project.

Data Collection

The data collection phase was an essential part of this project, serving as the cornerstone for developing and training the AI model. Since no pre-existing dataset was available at the start, a well-planned approach was necessary to obtain the required data. This involved leveraging both automated and manual methods to ensure a robust and comprehensive dataset was created, capable of supporting the objectives of the project.

Automated Data Crawling

Data crawling, an automated method of data collection, played a significant role in efficiently gathering a large volume of images. This approach used web-scraping techniques to pull images of vehicle interiors and dashboards from public sources, such as car-sharing forums, vehicle marketplaces, and automotive websites. The primary advantage of this method was speed—it allowed the collection of thousands of images within a short time frame, covering a wide range of car types, interiors, and dashboard designs. However, data crawling was not without its limitations. Many of the retrieved images were irrelevant, low-quality, or mislabeled, making it necessary to sift through and clean the data to ensure its usability. Despite this challenge, data crawling provided a diverse starting point for building the dataset.

Manual Data Searching

While automated crawling expedited the collection process, manual data searching was employed as a complementary method to ensure accuracy and relevance. This approach involved manually sourcing images by reviewing publicly available materials and selecting only those that met the project’s specific criteria. Each image was evaluated based on its quality, resolution, and relevance to the predefined categories, such as "clean interior," "dirty interior," and "dashboard faults." Although this process was labor-intensive and time-consuming, it added a layer of precision to the dataset, ensuring that only high-quality and meaningful images were included. The manual review process also helped in identifying gaps in the dataset that automated crawling might have overlooked, allowing for targeted data collection to fill those gaps.

Proprietary Data Contribution

In addition to externally sourced data, the company provided proprietary images from its operations, which were invaluable in enhancing the dataset. These images represented real-world scenarios and included both clean and dirty vehicle interiors as well as dashboards showing various operational conditions, such as engine and tire pressure faults. The inclusion of this proprietary data brought domain-specific insights into the dataset, making it more representative of the actual use cases. Due to a non-disclosure agreement (NDA) signed with the company, the specifics of these images cannot be detailed in this report. However, they added significant value to the project by improving the contextual relevance of the dataset and reducing the reliance on public sources.

Dataset Structuring and Categorization

Once the images were collected, the next step involved structuring the dataset into clearly defined classes. This was critical for model training, as each class needed to represent a specific condition or category to enable accurate predictions. For example, interior images were divided into clean and dirty subcategories, further classified based on specific areas such as the front left seat, rear right seat, armrest, or floor mat. Dashboard images were similarly organized into fault-free and fault-indicating classes, covering issues like engine warnings, tire pressure alerts, and oil level indicators. While some of these classes were pre-labeled in the proprietary data, others required manual annotation to ensure consistency and clarity across the dataset.

Challenges and Lessons Learned

The data collection phase was not without its challenges. Automated crawling often produced noisy data, requiring extensive cleaning to filter out irrelevant or redundant images. Manual searching, while effective, was resource-intensive and time-consuming. Balancing the efficiency of automated methods with the precision of manual efforts was a critical learning point during this phase. Additionally, the integration of proprietary data highlighted the importance of domain expertise in creating datasets that are both practical and relevant to real-world applications.

 

By combining automated crawling, manual searching, and proprietary data contributions, the team was able to build a diverse and balanced dataset. This multifaceted approach ensured the dataset was not only extensive but also representative of the scenarios the AI model would encounter during deployment. The meticulous organization and categorization of the data laid a strong foundation for the subsequent phases of preprocessing, training, and evaluation, ultimately contributing to the success of the project.

Data Preprocessing

Data preprocessing played a pivotal role in preparing the raw data for effective training and evaluation. This process ensured that the dataset was not only clean and well-organized but also aligned with the requirements of the AI model. Through a combination of cleansing, augmentation, and transformation, the dataset was optimized to improve the performance and generalizability of the final model.

Data Cleansing: Refining the Dataset

The first step in preprocessing was a thorough cleansing of the dataset to address errors, inconsistencies, and noise. This stage was critical in ensuring that the training data met a high-quality standard and that the model was trained on reliable and representative examples.

1. Correcting Misclassified Data
   • A significant portion of preprocessing involved rectifying misclassified images. Misclassifications could have arisen during the initial data collection or annotation phase, especially in the case of noisy or ambiguous images. Each image was reviewed and reassigned to the appropriate category based on its content. For example, if an image of a clean armrest was mistakenly labeled as dirty, it was reclassified to the correct "clean" class. This manual verification process, though time-consuming, was essential to avoid skewing the model’s training.
2. Handling Noisy Data
   • Some images within the dataset exhibited characteristics that could interfere with the model’s ability to learn effectively, such as excessive blurriness, poor resolution, or obscured content. These noisy images were categorized into a separate class, Other, to ensure that they did not influence the primary classification tasks. While these images were not used for training the current model, they were retained for potential future use, such as fine-tuning or exploring noise-resilient algorithms.
3. Excluding Irrelevant Data
   • Another challenge was the presence of irrelevant images that did not align with the defined project objectives. These included photos of unrelated objects or scenes mistakenly collected during the data collection phase. Such images were also moved to the Other class to maintain the dataset’s relevance and integrity. For example, an image of a vehicle’s exterior or a random roadside scene was excluded to focus solely on the vehicle's interior and dashboard.

Through this rigorous cleansing process, the dataset was refined into a more accurate and coherent structure, ensuring that each class contained high-quality, relevant examples.

Data Augmentation: Expanding the Dataset

For illustration purposes only, not an actual image

The second phase of preprocessing addressed the need for more diverse and extensive data. Deep learning models perform better when trained on large datasets with varied examples, as this allows them to generalize effectively to new and unseen inputs. However, collecting such large datasets is often resource-intensive and time-consuming. To overcome this challenge, data augmentation was employed.

1. Purpose and Benefits of Augmentation
   • Data augmentation is a technique used to artificially expand a dataset by applying transformations to existing images. These transformations simulate real-world variations, making the model more robust to changes in input conditions. In this project, augmentation was particularly valuable in addressing challenges such as:
     • Variations in photo angles, as users might capture images from different perspectives.
     • Differences in lighting conditions, which can affect image clarity and contrast.
     • Minor distortions caused by camera movements or focus.
2. Training Data Augmentation
   • For the training dataset, a wide range of augmentation techniques was applied to create new examples without altering the core characteristics of the images. These included: 
     • Flipping: Horizontally or vertically flipping the images to simulate different orientations.
     • Rotation: Rotating the images at various angles to mimic how users might position their phones while taking pictures.
     • Brightness Adjustment: Modifying the brightness levels to account for photos taken under varying lighting conditions.
     • Scaling and Cropping: Altering the scale of objects in the image or cropping parts of the image to simulate close-ups or wide views.
   • By incorporating these transformations, the training dataset was expanded significantly, increasing the model’s exposure to diverse scenarios and reducing the risk of overfitting.
3. Validation and Test Set Augmentation
   • Typically, validation and test datasets are left unchanged to preserve their real-world distribution. However, considering the specific challenges of this project—such as varying photo angles—limited augmentation was applied to these sets. Rotational transformations were used to ensure that the model could handle such variations during inference while maintaining the integrity of the evaluation process.

Through augmentation, the dataset size was increased by approximately one hundred times, enabling the model to learn more effectively and improving its robustness to real-world inputs.

Final Data Preparation for Model Input

Before the dataset was ready for training, additional preprocessing steps were applied to ensure compatibility with the AI model and consistency across the dataset:

1. Resizing Images
   • All images were resized to match the input dimensions required by the pre-trained model used for transfer learning. This step ensured that the images fit seamlessly into the model’s architecture without requiring further resizing during training.
2. Standardization
   • Pixel values were standardized to maintain uniformity across the dataset. This helped reduce the impact of variations in brightness and contrast, enabling the model to focus on the content of the images rather than extraneous factors.
3. Normalization
   • To align with the pre-trained model’s original training conditions, pixel values were normalized to a specific range, such as 0 to 1. This step not only improved the convergence of the model during training but also ensured that the inputs were consistent with the model’s expected format.

Summary of Preprocessing Steps

By combining meticulous data cleansing with extensive augmentation and careful preparation, the dataset was transformed into a high-quality resource for training the AI model. These preprocessing efforts laid the foundation for a robust and reliable system capable of accurately identifying vehicle interior cleanliness and dashboard faults. The resulting dataset not only met the project’s immediate objectives but also provided a scalable framework for future enhancements.