Abstract

Object detection has significantly advanced with deep learning but faces challenges under image distortions like noise, compression, blur, fog, and snow. This issue is critical in applications such as self-driving cars and healthcare. While defence methods aim to enhance robustness under distortions, maintaining performance on clean images remains a challenge.
To address this challenge, we propose a novel defence approach that generates copies of the original training images and adds distortion-like content to these copies at the pixel level. We balance the number of distorted pixels to prevent bias during learning. Our approach includes two augmentation methods to generate augmented samples: AISbod, which uses artificial immune systems (AIS), and GSES, which employs kernel density estimation (KDE).

Our AISbod uses AIS to distort the original sample (antigen) through cycles of “select, clone, mutate, select” until the augmented data (antibody) reaches a specified similarity to the antigen. However, AIS is limited in diversifying generated antibodies and is computationally expensive. Therefore, in GSES, we create samples by selecting pixels from the original samples, estimating the pixel distribution from multiple distorted versions of the original samples via KDE, and then replacing the selected pixels with new values sampled from the estimated distribution. GSES generates more diverse data than AISbod.

We evaluate our methods on 15 image distortions using state-of-the-art object detection models like DINO and YOLOv7. Our methods improve accuracy under distorted and clean images and remain consistent across datasets and detection models. For instance, DINO on the COCO dataset shows a 4.50% improvement under clean samples, 8.40% on average across all distortions, 2.50% under snow, and 29.30% under impulse noise. The observed improvement is due to the weight regularization effect of our methods, which is evident in the smoother convergence of training and validation loss curves, indicating reduced learning fluctuations and a more stable optimization path. Additionally, the narrower gap between the curves suggests reduced over-fitting, leading to better generalization to unseen data. Simulations indicate that our approach outperforms related defence methods against distortions and extends beyond object detection, improving accuracy in image classification and object tracking models.