Learned Image Correction for Object Detection

An RL agent that repairs a degraded image before a frozen detector sees it

This project was conducted at IIT Kanpur, India, under the Ministry of Electronics and Information Technology (MeitY), Government of India. The goal was to make learning-based perception hold up under the degraded imaging conditions robots meet in the field.

Object detectors degrade sharply under noise, blur and low light, and the usual remedy — denoising or enhancing the image so it looks better to a human — does not reliably help the detector. The corrections a network needs are not always the ones a person would choose.

This project takes the detector’s own performance as the training signal. A PPO agent applies a correction to the incoming image, a frozen YOLOv3 detector runs on the result, and the detection quality becomes the reward. The policy therefore optimises directly against downstream detection rather than against a hand-designed image-quality metric.

Approach

  • Frozen detector. YOLOv3 stays fixed, so improvements come from the correction policy rather than from fine-tuning the perception network.
  • Reward from detection. Confidence, IoU and NMS thresholds feed the reward, tying the policy to the metric that actually matters.
  • Policy variants. A convolutional policy, a ResNet-backbone variant, and an intrinsic-curiosity variant for sparse-reward regimes.
  • Dataset. PASCAL VOC, with annotations rescaled to the working resolution.

Code

Released at gaurav-gaurav/RL-for-object-detection — training and evaluation scripts, policy variants and the annotation tooling.