WordDetectorNet Pixel Regression

WordDetectorNet is a model that adopts per-pixel bounding-box regression instead of conventional anchor-based detection and NMS (Non-Maximum Suppression) methods.

How It Works Every pixel that the network classifies as a 'word pixel' regresses 4 scalar values representing the distances from that pixel to the top/bottom/left/right of the bounding box. Through this, each pixel generates one candidate box, resulting in thousands of overlapping candidate boxes per word.

The candidates are then merged via DBSCAN clustering. Using $1 - \text{IoU}$ as the distance metric, spatially overlapping candidates are grouped together, and the median box of each cluster is adopted as the final detection result.

Architecture and Features

• Structure: Uses a ResNet18 backbone with an FPN-style decoder, and the final head outputs 6 channels per pixel (2 segmentation logits + 4 distance values).
• Advantage: There is no need to separately tune hyperparameters such as anchor sizes or NMS thresholds.
• Disadvantage: Computing the IoU matrix between candidate boxes has $O(n^2)$ complexity, which becomes a bottleneck for runtime speed, and DBSCAN's parameters must be set manually, making end-to-end training difficult.