Hospitality Sector

Hotel Image Recognition and Classification Solution

Custom Backend Image Recognition

Developing a model that can categorize hotel images?
Challenge Accepted!

Hotel Image Recognition and Classification Solution

Process & Story

The process of image recognition is always challenging, but it is challenging when dealing with images that are not well-labeled or contain a variety of objects. In the context of hospitality, this means trying to develop a model that can correctly identify and categorize photos of hotels. This was the task at hand for our team as we sought to create an AI solution based on visual image recognition.

Problem

Hospitality platforms deal with hundreds of thousands of images for hotel offers. For each offer, booking engines present several selected images that vary from the bedroom, bathrooms, lobby, pool, restaurant, window view, and nearby attractions. To provide the best user experience, the platform should show an optimal number of photos from each category to present the hotel offer in the best possible way visually. We’ve observed that booking engine sites deal with vast numbers of hotel photos, some with very low image quality or not particularly relevant to travelers, but the problem here is the lack of description for content in the image resulting in the wrong order of the photos.

To show the right image to the user, the platform must first categorize the image. This can be done either manually by visual content specialists or automatically with the help of machine learning and visual image recognition.

It is impossible to keep up with a costly manual process for big hospitality platforms that integrate with various vendors and update their offers daily. Booking engines are in great need of an artificial intelligence tool that can automatically detect and categorize photos for all their hotel offers.

This process of image classification is a complex task as it not only requires identifying what’s in the image but also understanding the context of the image and how it relates to other images in the same category.

Solution

The solution was to create a machine learning model that could sift through the image data and automatically categorize the photos. The first step was to gather a dataset of images representing the different types of hotels and their interiors found in the hospitality industry. This dataset was used to train a machine learning model that could be used for image recognition.

As this is an image recognition problem, we decided to apply ResNet architecture for the machine learning model. The Residual Network (ResNet) is a type of Convolutional Neural Network (CNN) architecture that addresses the “vanishing gradient” issue, allowing for networks with hundreds of convolutional layers to outperform shallower ones. The ResNets architecture and convolutional neural networks are frequently used in deep learning computer vision applications such as object detection, image classification, and image segmentation.

For the deep learning model building and training, we used the fast.ai library, which increases the level of abstraction of PyTorch. It is relatively new but already supports good practices and is always up to date with advancements in deep learning. We fine-tuned the ResNet deep learning model pre-trained on the ImageNet dataset with the hotel images dataset.

Once the model was trained, it was then tested on a new set of images to see how well it could categorize them. The results were promising, and the model was able to identify and categorize a variety of hotel images correctly. However, there were still some challenges that needed to be addressed.

The next step was to create a prototype application that would allow users to upload photos and identify the category of each photo or to do the same through our dedicated API endpoints.

Case Study Schema Hotel Image Recognition and Classification Solution

Tools

Google Cloud Platform

PyTorch

Python

Have a similar project in mind?

LET'S TALK!

Challenges

Gathering data and building a hotel images training set
We needed to create a dataset for the training process from scratch, and we decided to rely on free and open-source resources. Indoor scene recognition from Kaggle and MIT resources datasets fitted pretty well for our task. They needed some preprocessing, and there were many more categories with unrelated images for hotel recognition. When it comes to training data, it is always easier when you have more than required, as you can exclude unnecessary categories easily. The second batch of hotel images was imported by simply using Google Graphics. They have an open API, and the GitHub repository provides a convenient library.

Limited data set and data augmentation
Image training set data is a never-ending stream. When it comes to image recognition tasks, more data means better accuracy. However, sometimes gathering more data is not possible or practical. In those cases, you can resort to data augmentation, which is the process of modifying existing data to increase its size and diversity. This is precisely what we did with our image data. We randomly rotated, flipped, and scaled the images to make them look different. This not only helped us increase the size of our dataset but also improved the accuracy of our model. We ended up with a large-scale curated dataset of images for hotel recognition.

Preprocessing data for image classification
Most computer vision models are susceptible to image noise and other artifacts. This can be a significant challenge when you are working with a limited amount of data. In our case, we used a preprocessing step to clean up the hotel room images and make them more suitable for image recognition tasks. We manually verified the types and number of available categories for classification.

Transfer learning
When it comes to deep learning models, you usually start with a pre-trained model and then tweak it for your specific task. This is known as transfer learning. It can be a helpful strategy when you don't have enough data or when your data is not suitable for training a model from scratch. Only a few top layers of weights are responsible for choosing the proper class. All layers below recognize elements with a smaller level of abstraction, i.e., gradients, curves, lines, pixels, etc. So we only have to retrain the top layers from scratch and slightly adjust the others. We used a pre-trained deep learning model initially trained on the ImageNet dataset. This saved us a lot of time and resources, resulting in good accuracy. You can read more about our approach to transfer learning here.

Confusing categories from test set
During model training, we identified several shortcomings. Our model struggled with correctly classifying some of the categories like front hotel vs. outdoor, reception vs. lobby, or bar vs. restaurant. Some bars may be confused with restaurants, as well as a lobby and reception. Another challenge was that the model sometimes had trouble distinguishing between a hotel image and an image of a nearby attraction. This is because some hotels are located close to popular tourist destinations, and it cannot be easy to distinguish which picture is from the hotel and which image is from the attraction. To overcome this, we worked on a better category setup and removing unnecessary classes.

Our other projects: