Extruding Resin Direction Angle Inference Pipeline

---

1. Defined segment of resin for semantic segmentation.

Sequence Resin Data

Segmentation Label

2. Calculation of angle about extruding resin direction

Eigenvector extraction using 2D point coordinate data of the line segment area. The angle calculation uses the atan function.

3. Proposal Semantic Segmentation Model Archtecture

Backbone: Deeplabv3 (resnet101, Atrous Spatial Pyramid Pooling) Loss Function: Boundary Focal Loss (proposal method) coco dataset Image Pre-Processing

4. Boundary Focal Loss (Novelty)

English ver: Designation of a boundary area focus loss function that reflects a higher loss value to a sample with high uncertainty with a low standard deviation by specifying the boundary area adjacent to an object as the standard deviation of softmax logits and applying a weight calculation function that is inversely proportional to the standard deviation. Korean ver: 객체와 객체 사이에 인접한 경계 영역을 softmax logits의 표준편차로 특정하고 표준편차에 반비례하는 가중치 산정함수를 적용하여 표준편차가 낮은 불확실성이 높은 표본에 더 높은 손실 값을 반영하는 경계 영역 초점 손실함수 고안.

5. mIoU Metric result compare with other loss function

Cross Balancing Loss, Cross Entropy, Focal Loss, Boundary Focal Loss result box plot.

외곽선분 영역 Left, Right Line 분할 결과 비교 그래프.

Focal Loss와 Boundary Focal Loss로 추론 결과를 정답 데이터를 기반으로 한 압출 수지 진행 방향 각도와 MAE 오차 비교 결과

6. Inference Result Image

입력데이터, Segmentation 추론 결과, 방향 각도 계산 결과