This project was created to count the number of coins and estimate the total value. With a few image of coins, we will agruementate small dataset and try to using items detection network to approach our goal. Swiss coins dataset is the only that at our disposal on internet, so we altimately choose this as our target sets.
This is a two-stage method.
At first, Hough Method involved, however, the methods is parameter-sensitive. so we alternatively find the edge and counter of the images, and find a boundingbox for each item. With the help of the box, we masked those small items, coins, from the big image. As we know, coins have its unique features, using this, we make a filter to extract those items that have a higher posiblity to be coins. Then we can figure out the number of the coins.
Secondly, we leverage a lite network, mobilenet, to classify the extracted small images. With the help of deeplearning, we can finally count out the total value of the showed image.
- still background sensitive.
- terrible performance
- not end to end
NOTE: worth to mention, a part of my code got reference from https://www.kaggle.com/kmader/mobilenet-classification.
We will try a new approach, utilizing YOLOv5 serial networks. Detection networks provid an end to end framework, which has an absolute better performance.
UPDATE: With the help of the existed project https://github.com/earthsaharat/Thai-coin-detection, we edited part of its code, and after our exam, this programme provides a more robust result compared with the former.
Edge Detected:
Hough Circle: At very beginning, we want to use this method to extract the coins, but gradually I find that this algorithm excessively rely on the quality of the edge detection. Apart from this, controlling the input parameter also make this stratage unrobust. So, I give up this pipeline.
Detected Bondingboxes:
Delect Using NMS and priori knowledge:
