DBSCAN-Based WordDetectorNet Architecture Analysis

WordDetectorNet proposes a unique architecture that combines pixel-level bounding box regression with DBSCAN clustering, instead of the conventional anchor-based detection and NMS (Non-Maximum Suppression) approach.

Core Mechanism

• Every pixel classified by the network as a 'word pixel' also regresses 4 distance values (top/right/bottom/left) for the bounding box containing that pixel.
• This generates thousands of overlapping candidate boxes, which are clustered using DBSCAN with the metric distance = 1 − IoU.
• Finally, the median box of each cluster is adopted as the detection result.

Architecture Composition

• Backbone: ResNet18 (modified for 1-channel grayscale input and intermediate feature extraction).
• Decoder: An FPN-style decoder that upscales and combines features from all scales.
• Head: Outputs 6 channels per pixel (2 segmentation logits + 4 distance values).
• Loss: Trained with cross-entropy and IoU loss weighted equally.

Pros and Cons

• Pros: No need for anchor configuration or NMS threshold tuning, and the IoU-based clustering approach is conceptually clear.
• Cons: Computing IoU for pairs of candidate boxes has $O(n^2)$ complexity, becoming a bottleneck for execution speed, and DBSCAN's eps hyperparameter must be set manually, making end-to-end training difficult.