Semi-automatic Segmentation & Alignment of Handwritten

No. segmentation boxes: The total number of boxes after segmentation and be- fore alignment.

• 
  Segmentation boxes error (EsB): The error between the number of boxes after segmentation and the number of words in the GT. Calculated according to Equa- tion 2.
  

• 
  No. alignment boxes: The total number of boxes after alignment.
  

• 
  Alignment boxes error (EaB): The error between the number of boxes after alignment and the number of words in the GT. Calculated according to Equation 4.
  

Error = |1 −
Aligned_words
GT_W ords ^{| (4)}

• 
  No. words: The total number of aligned words for one image.
  

• 
  Words error (EW): The error between the number of words after alignment and the number of words in the GT. Calculated according to Equation 3.
  

• 
  No. lines: The total number of lines after segmentation.
  

• 
  Lines error (EL): The error between the number of lines after segmentation and the number of lines in the GT. Calculated in the same way as in Equation 2 and Equation 3.
  

6. Results

The following chapter consists of two sub-chapters. The first part presents the results from the performance evaluation experiments. The second part visualises the resulting algorithm pipeline.

1. Performance of the algorithm

When comparing the results of the algorithm’s performance (Table 3, 4, 5) on both data sets it is evident that it performs worse on Labour’s Memory in terms of IoU. Although IoU is a crucial performance metric, it does not provide any information on how many boxes are subject to segmentation faults, or how many words from the GT are missed in the alignment. For more information, metrics such as segmentation box error, alignment box error, and word error need to be evaluated. A combination of the calculated metrics will be able to give a better representation of the overall performance. See Supplemen- tary Files to access the raw data used in the performance evaluation.

1.