data/config-Y-T-G.yaml

train: sam_40_mss_lines/train/images
val: sam_40_mss_lines/val/images
nc: 1
names: ['text_line']

scales: # model compound scaling constants, i.e. 'model=yolo11n-seg.yaml' will call yolo11-seg.yaml with scale 'n'
  # [depth, width, max_channels]
  n: [0.50, 0.25, 1024]
  s: [0.50, 0.50, 1024]
  m: [0.50, 1.00, 512]
  l: [1.00, 1.00, 512]
  x: [1.00, 1.50, 512]

# Backbone (P3-P6)
backbone:
  # First layers, using convolutions and C3 blocks for feature extraction
  - [-1, 1, Conv, [64, 3, 2]]  # P1/2
  - [-1, 1, Conv, [128, 3, 2]]  # P2/4
  - [-1, 2, C3k2, [128, False, 0.25]]  # C3 block for P2/4

  - [-1, 1, Conv, [256, 3, 2]]  # P3/8
  - [-1, 2, C3k2, [256, False, 0.25]]  # C3 block for P3/8

  - [-1, 1, Conv, [512, 3, 2]]  # P4/16
  - [-1, 2, C3k2, [512, True]]  # C3 block for P4/16

  - [-1, 1, Conv, [1024, 3, 2]]  # P5/32
  - [-1, 2, C3k2, [1024, True]]  # C3 block for P5/32

  - [-1, 1, Conv, [1024, 3, 2]]  # P6/64
  - [-1, 2, C3k2, [1024, True]]  # C3 block for P6/64

  - [-1, 1, SPPF, [1024, 5]]  # SPPF block for improved feature extraction
  - [-1, 1, C2PSA, [1024]]  # Cross-stage partial aggregation (PSA)

# Segmentation head for predicting masks
head:
  - [-1, 1, nn.Upsample, [None, 2, "nearest"]]  # Up-sample P6 (nearest neighbor)
  - [[-1, 8], 1, Concat, [1]]  # Concatenate backbone P5 features
  - [-1, 2, C3k2, [1024, False]]  # C3 block for P5

  - [-1, 1, nn.Upsample, [None, 2, "nearest"]]  # Up-sample P5
  - [[-1, 6], 1, Concat, [1]]  # Concatenate backbone P4 features
  - [-1, 2, C3k2, [1024, False]]  # C3 block for P4

  - [-1, 1, nn.Upsample, [None, 2, "nearest"]]  # Up-sample P4
  - [[-1, 4], 1, Concat, [1]]  # Concatenate backbone P3 features
  - [-1, 2, C3k2, [512, False]]  # C3 block for P3

  # Additional convolutional layers for refinement
  - [-1, 1, Conv, [512, 3, 2]]  # Conv on P4
  - [[-1, 18], 1, Concat, [1]]  # Concatenate refined P4 features
  - [-1, 2, C3k2, [1024, False]]  # Further refinement on P4 features

  - [-1, 1, Conv, [1024, 3, 2]]  # Conv on P5
  - [[-1, 15], 1, Concat, [1]]  # Concatenate refined P5 features
  - [-1, 2, C3k2, [1024, True]]  # Further refinement on P5 features

  - [-1, 1, Conv, [1024, 3, 2]]  # Conv on P6
  - [[-1, 12], 1, Concat, [1]]  # Concatenate refined P6 features
  - [-1, 2, C3k2, [1024, True]]  # Further refinement on P6 features

  # Final segmentation output (mask output layer)
  - [[21, 24, 27, 30], 1, Segment, [nc, 32, 256]]  # Segmentation head