read
Deliverables
- Extending texted detction to detect text from other parts of the video.
The main challenge while implementing this part was to identify pieces of text that belong to the same sentence. In the previous version of textdetection that I used for detecting parts of the ticker, I had limited the search to the bottom of the video. Since there was only a single line of text, I simply had to recrd the lowest and hightest y and x coordinates (xmin,ymin and xmax,ymax) and crop out the text regions using cropped = frame[ymin:ymax,xmin:xmax].
In order to identify pieces of detected text that belong to the same sentence and crop it out as a whole, I implemented the following procedure:
Input: Set of vertices returned after implementing EAST
Output: Sets of continuous text
1. for each frame:
vertices = run_east_text_detection_algorithm()
''' dictionary keys = xmin for each box
values = x0,y0,c1,y1 '''
final_boxes =dict()
1. for v in vertices:
1. if v.x0 is in final_boxes:
key = v.x0
if final_boxes[key].x0 > v.x0: final_boxes[key].x0 = v.x0
if final_boxes[key].x1 < v.x1: final_boxes[key].x1 = v.x1
if final_boxes[key].y0 > v.y0: final_boxes[key].x0 = v.y0
if final_boxes[key].y1 > v.y1: final_boxes[key].y1 = v.y1
else
add v to final_boxes
2. for v in final_boxes:
1. cropped = frame[ymin:ymax,xmin:xmax]
2. save(cropped)