To get correct results

[Kaung_Khant_Zaw]

Hi… I’m implementing the project which will tell the correct labels. To get the correct result, what should i do. Now I annotated about 90 images for each classes and test it. But the results are still not correct. Sometimes correct and sometimes not.

[Lenny]

Based on the information provided in the extracts, there are several steps you can take to improve the accuracy of your model:

1. Increase the number of images: The more images you have in your dataset, the better your model will be able to learn and generalize. It’s recommended to collect a few hundred images in your dataset before starting to measure data drift source.

2. Improve your annotations: The quality of your annotations can significantly impact the performance of your model. Make sure to draw accurate and tight bounding boxes around your objects of interest. Consistency in annotation is also important. You can refer to the Roboflow annotation guide and 7 annotation tips for best practices.

3. Use Roboflow’s Label Assist: Roboflow Annotate features Label Assist, an automatic annotation tool that can help you annotate images more accurately and efficiently.

4. Check for mislabeled images: Mislabeled images can negatively impact your model’s performance. Roboflow provides tools to help you identify and correct mislabeled images in your dataset source.

5. Apply data augmentation: Data augmentation can help your model generalize better by providing it with more varied examples during training. Roboflow supports a variety of augmentation techniques source.

Remember, machine learning models improve with more data and better quality data. So, keep iterating on your dataset and model training process.

---

Hi! I’m Roboflow’s AI powered bot. This is an automated response powered using AI with knowledge from our docs, blog, and forum. If this solves your issue, please mark it as the solution using the checkbox icon. If this doesn’t solve your issue, please follow up with what you’ve tried/why, and the community will continue to chime in as usual.

[Kaung_Khant_Zaw]

When creating the project, i choose object detecting… should i choose another type for this kinds of class differentiation.

[Jacob_Witt]

Hi @Kaung_Khant_Zaw - are the class names 200, 500, and 1000?

Couple things you can try:

1. Do object detection and draw a bounding box around just the text.
2. Try a 2-stage model where you identify coins first, and then pass that input through OCR like DocTR.
3. Just add more images - my guess is this will start to work fine around ~300 images.
4. Make sure you aren’t adding too many augmentations or preprocessing steps; if the images are too distorted it can hurt model performance.

[Kaung_Khant_Zaw]

In the project, there will be so many classes. Both (200,500,1000) and (200,500,1000) in another languages. In the image, that is only one class meaning, there will be another brand bottle cap with (200,500,1000) will come. And the text will come various position like rotating and so on. So, if I use the OCR, the ocr must read both the eng and other text in various. I have try many ocr including DocTR but in that test, some are not detecting the text and some are detect but wrong. So I don’t think using the ocr will get the correct result. So what is the best way to get the correct result? Thank you for the reply.

[Kaung_Khant_Zaw]

Like this photo. The text will be various position and in the myn language. So can it identify the correct classes only with the adding more data and training the data?

The classes will be like

in this image
1.brand1_200
2.brand1_500
3.brand1_1000

in the above image the classes will be like

1. brand2_200
   2.brand2_500
   3.brand2_1000

there will be about 50 classes in the project. So the classes should be identify correctly. And one more thing is in the brand, the difference will be only the number and the design will be the same.

[Jacob_Witt]

How does OCR work if you only crop the bottlecaps? That’s my suggestion- have a generic bottle-cap detector, and then run OCR / classification on the cutouts.

[Kaung_Khant_Zaw]

Thank you so much for the reply…

The problem is some of the text are detect as wrong even in the google lens or others ocr… so using OCR will have problems for the future…

So do you mean like I have to train the model only for detecting the bottle cap and then crop the bottlecaps and use the other model which is trained for the design classes? So for the designs to train, I have to use the classification type project?

Thank you so much for your time…