Neural Style Transfer for Improved Object Detection

Read my write-up!

Summary

Using neural style transfer (NST) to generate additional training samples can improve the generalizability of convolutional neural networks (CNNs) for object detection tasks.

Dataset

• Base images from Imagenette (subset of ImageNet)
• 10 object classes: chainsaw, gas pump, tench, french_horn, church, english_springer, golf ball, garbage truck, parachute, cassette player
• Source: https://github.com/fastai/imagenette

Methodology

1. Neural Style Transfer
   • Used TensorflowHub's Arbitrary Image Stylization model
   • Applied 5 distinct artist styles from Best Artworks of All Time
   • Reference: https://www.tensorflow.org/hub/tutorials/tf2_arbitrary_image_stylization

Example Images

Results

Training a CNN with the augmented dataset (original + stylized images) yielded roughly 15% improvement for accuracy, precision, and recall.

These results suggest that neural style transfer is an effective technique for data augmentation in object detection tasks.