Dual-Branch Scene Text Detection Model Development

Current Status:
I already have a fully runnable MMDetection/YOLOv3-based framework with training and evaluation scripts, core source code, and custom modules for the Tampered_IC13 dataset.
The dataset is prepared and partially configured for training.
The framework can already run standard YOLOv3 object detection training and evaluation.

Requirements:
Implement a dual-branch architecture with RGB and frequency-domain branches.
Develop a fusion module to effectively combine features from both branches.
Train and evaluate the model on the Tampered_IC13 dataset (and potentially other datasets) to achieve a target F1 score ≥ 0.85, with balanced Precision and Recall.

Deliverables include:
Trained model weights (.pth)
Complete training and evaluation logs (including hyperparameters and loss curves)
Detailed evaluation metrics (Precision, Recall, F1)
A reproducibility guide (instructions to run and replicate the results in my environment)
Optimize the model’s performance to exceed baseline results.

Required Skills:
Proficiency in PyTorch and MMDetection/YOLOv3
Experience with frequency-domain image processing (e.g., FFT, DCT)
Strong background in object detection model design and optimization