Improved Road Pothole Detection and Instance Segmentation Using Mask RCNN with Histogram Equalization and Adam Optimization
This study proposes an improved Mask R-CNN framework enhanced by histogram equalization and the Adam optimizer to achieve a 90.4% mAP in road pothole detection and instance segmentation, significantly outperforming the baseline model for intelligent road monitoring.
Original paper licensed under CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer
Imagine you are a detective trying to find a specific type of clue hidden in a messy, chaotic room. In the world of computer science, this "room" is a digital image, and the "clue" is a specific object, like a pothole on a road. For a computer to act like a detective, it needs to be trained using a special kind of brain called a "Deep Learning" model. Think of these models as students who learn by looking at thousands of examples. One popular type of student is called Mask R-CNN. While a regular student might just point and say, "There's a pothole!" and draw a square box around it, Mask R-CNN is the overachiever that also traces the exact, jagged outline of the pothole, pixel by pixel. This is called instance segmentation. But here's the catch: if the room is dark, the walls are painted weird colors, or the lighting is flickering, even the best student gets confused. This is why researchers care about image processing—it's like giving the detective better glasses or a flashlight so they can see the clues clearly, no matter how messy the environment is.
The researchers in this paper, Chuan-Bi Lin and Alok Kumar Sharma, decided to upgrade their "detective" to make it even better at finding road potholes. They knew that standard training methods sometimes struggle when the road images are dark, blurry, or have uneven lighting. To fix this, they introduced two main upgrades to their system. First, they used a technique called Histogram Equalization. Imagine you have a photo that looks washed out and gray; this technique is like turning up the contrast on a TV screen. It stretches out the colors and shadows so that the dark holes in the road pop out clearly against the lighter pavement, making them much easier for the computer to spot.
Second, they changed how the computer "learns" from its mistakes. Usually, these models use a method called Stochastic Gradient Descent (SGD), which is a bit like a hiker trying to find the bottom of a valley by taking small, random steps. Sometimes, the hiker gets stuck in a small dip and thinks they've reached the bottom. The researchers swapped this for a smarter hiker using the Adam optimizer. Adam is like a hiker who remembers where they've been and adjusts their steps dynamically, helping them find the true bottom of the valley (the best possible model) much faster and more reliably.
The team tested their improved system on a collection of 335 road images they found online. Because that wasn't enough to train a super-smart detective, they used a trick called data augmentation. They took the existing photos and created new ones by rotating them and flipping them horizontally, effectively turning 335 images into over 1,800 training examples. They also made sure to draw the "ground truth" (the correct answer) using precise polygon shapes that followed the weird, irregular edges of the potholes, rather than just simple boxes.
When they ran the experiments, the results were quite promising. The original, un-upgraded version of the model managed to find potholes with a score called Mean Average Precision (mAP) of 0.779. However, after adding the contrast-boosting flashlight (Histogram Equalization) and the smarter learning method (Adam optimizer), the model's score jumped to 0.904. This means the new system was significantly better at both finding the potholes and drawing their exact shapes. The researchers found that the biggest boost came from the image enhancement, which helped the model "see" the potholes much more clearly, especially in tricky lighting.
They also compared their upgraded detective against other famous models like YOLOv2, SSD300, and Faster R-CNN. The proposed Mask R-CNN framework came out on top with its 0.904 score, beating the next best competitor (Faster R-CNN) which scored 0.732. The paper suggests that combining better image preparation with smarter training algorithms makes a huge difference. While the system isn't perfect—it still had a few minor hiccups with very irregular shapes or extreme lighting—the study concludes that these improvements make the technology much more reliable for keeping roads safe. The authors suggest that in the future, this kind of system could be part of a larger network to monitor road conditions in real-time, helping to fix roads before they become dangerous.
Drowning in papers in your field?
Get daily digests of the most novel papers matching your research keywords — with technical summaries, in your language.