Vision-Based Calorie Estimation for Bangladeshi Street Food: A Comparative Study of Detection and Regression Models

Obesity is a growing health challenge worldwide, affecting hundreds of millions of people and straining healthcare systems. A key part of managing weight is understanding how many calories are in the food we eat. For decades, scientists have tried to build computer programs that can look at a photograph of a meal and calculate its energy content. These systems rely on a branch of artificial intelligence called computer vision, which allows machines to "see" and identify objects in images much like humans do. However, most of these existing tools were trained on Western dishes, such as burgers or salads, where portion sizes are often standard or the food sits on a plate with clear boundaries. They struggle significantly with the vibrant, varied, and often irregular street foods found in other parts of the world, where a single serving might be a fried ball, a flaky pastry, or a stack of flatbread, and where the size of the portion can vary wildly from vendor to vendor. Without accurate data on these specific foods, people in regions like Bangladesh cannot easily track their daily intake, leaving them without a crucial tool for making healthier choices.

A team of researchers from Premier University in Chattogram, Bangladesh, set out to solve this specific problem by creating a system designed exclusively for the street foods of their country. They recognized that simply recognizing the type of food was not enough; the computer also needed to know exactly how big the piece was to estimate the calories correctly. To achieve this, they developed a method that uses a common object found in every pocket—a five Taka coin—as a built-in ruler. By placing this coin next to the food in a photograph, the system can calculate the real-world size of the snack, regardless of how close or far away the camera was held. This approach bypasses the need for complex 3D sensors or strict photo-taking rules, making the technology practical for everyday use on a standard smartphone.

The researchers began by building a new library of images, a dataset that did not exist before. They collected nearly four thousand photographs of six popular street foods: Singara, Somusa, Puri, Peaju, Beguni, and the reference coin itself. These images were taken in various conditions, from professional studio lighting to the chaotic, dimly lit environments of actual street stalls, ensuring the system would work in the real world. They then tested five different types of advanced computer vision models to see which one could best identify and outline the shape of each food item. These models act like digital eyes, drawing a precise border around every piece of food in the image. After comparing their performance, the team found that one specific model, known as YOLO11n, was the most accurate. It successfully identified the food and its boundaries with a high degree of reliability, even when the items were clustered together or partially hidden.

Once the computer had identified the food and drawn its outline, the next step was to measure it. The system detected the five Taka coin in the image, which has a known diameter of 25.5 millimeters. By comparing the size of the coin in pixels to the size of the food in pixels, the software calculated a scaling factor. This allowed it to translate the digital image into real-world dimensions, determining the actual height, width, and surface area of the snack. With these physical measurements in hand, the researchers fed the data into a separate set of mathematical tools designed to find patterns and make predictions. They tested several different prediction engines to see which one could best guess the calorie count based on the size and type of the food. The results showed that a method called Random Forest, which works by combining the decisions of many simple decision-makers to reach a single, robust conclusion, produced the most accurate estimates.

The final system proved to be highly effective. When tested, the combination of the best vision model and the best prediction engine could estimate the calorie content of these street foods with a very small margin of error. The average difference between the estimated calories and the actual calculated value was less than six calories, and the system explained 95 percent of the variations in the data. This level of accuracy suggests that the method is reliable enough to be used in mobile applications for dietary monitoring. The researchers also noted that their system is efficient enough to run on standard hardware, avoiding the heavy computational costs that often make such technology impractical for everyday use. By focusing on local cuisine and using a simple, accessible reference object, this work fills a significant gap in nutritional technology, offering a practical way for people in Bangladesh to understand the energy content of the food they eat every day.