This repository contains the code I used for the Tampered Text in Images (TTI) Tianchi Competition. The goal of the competition is to develop models that can accurately detect and locate tampered text in images.
The TTI dataset contains 19,000 text images, with 15,994 of them manipulated using various techniques. Each image is annotated with a binary mask indicating the tampered location. To preprocess the data, I resized the images to 256x256 pixels and normalized the pixel values to between 0 and 1.
For this competition, I experimented with various deep learning models and ultimately settled on using the EfficientNet-B4 architecture. EfficientNet-B4 is a powerful convolutional neural network that has achieved state-of-the-art results on many computer vision tasks.
To further improve the performance of the model, I employed transfer learning techniques by fine-tuning the pre-trained weights of the model on the TTI dataset. Additionally, I used data augmentation techniques such as random rotations, flips, and zooms to increase the variability of the training data and help prevent overfitting.
My model achieved a score of 81.18 on the competition leaderboard, which was in the 68/1256 of all submissions.