I have a dataset of 1000 labeled images of CAPTCHA from my college website, and I want to train a model that can accurately solve similar CAPTCHAs on unseen data. The CAPTCHA typically consists of 6 alphanumeric characters (A-Z, 0-9). Despite trying several approaches, the model fails to achieve high accuracy.

Dataset:

• Number of Samples: 1000 labeled images.
• CAPTCHA Type: 6 alphanumeric characters.
• Image Example: [Attach an example CAPTCHA image].
• What I’ve Tried:
  Convolutional Neural Network (CNN):

I created a CNN model with multiple convolutional and dense layers.
Flattened the output to feed into separate dense layers for each character.
Achieved poor generalization on unseen CAPTCHAs.
Transfer Learning:

Used MobileNetV2 as the feature extractor with a custom head.
Adapted input size (grayscale to RGB conversion).
The model was prone to overfitting due to limited data

I want to train a model that:

• Can handle unseen CAPTCHAs with high accuracy.
• Effectively decodes sequences of alphanumeric characters.

What is the best approach or architecture for solving CAPTCHA like this? Should I use:

• CNN for feature extraction followed by LSTM for sequence decoding?
• A fully convolutional architecture (e.g., CRNN)?
• Any other approach that works better for limited datasets?

If possible, could you suggest a complete model architecture, preprocessing steps, or loss function setup that might help? Any tips for augmenting a small dataset for this task?