LEARNet: A Learning Entropy-Aware Representation Network for Educational Video Understanding

. 2025 Dec 19;28(1):3. doi: 10.3390/e28010003

Algorithm 1: Temporal Information Bottleneck (TIB)

1:
Input: Sequence of video frames F = {f₁, f₂, …, f_n} frame rate fps
2:
Output: Pedagogically significant keyframes K = {k₁, k₂, …, k_r}, moderate significance frames K’ = {k’₁, k’₂, …, k’_s}
3:
// Step 1: Pedagogical Information Scoring
4:
for i ← 2 to n do
5:
Compute MSEᵢ ← $\frac{1}{N} \sum_{k = 1}^{N} {(I_{i} (k) - I_{i - 1} (k))}^{2}$ // structural saliency
6:
Compute SaliencyMapᵢ ← ComputeSaliencyMap(fᵢ)
7:
Compute SaliencyValueᵢ ← Mean (SaliencyMapᵢ) // content saliency
8:
Set weight α ← 0.67
9:
Compute SignificanceScoreᵢ ← α × MSEᵢ + (1 − α) × SaliencyValue_i//pedagogical information score
10:
end for
11:
//Step 2: Instructional Transition Identification

Apply Clusters ← DensityPeakClustering({SignificanceScore₂, …, SignificanceScore_n})

12:
Initialize candidate keyframes: K ← { $f_{c}$ | $f_{c}$ ∈ Clusters.Centroids}
13:
// Step 2: Domain-informed pruning
14:
Initialize K’ ←∅
15:
for each frame k∈ K do
16:
Compute duration ← ComputeFrameDuration(k, fps)
17:
Compute personArea ← ComputePersonArea(k)
18:
Compute ssim ← ComputeSSIM (k, previousKeyframe)
19:
Compute saliencyScore ← Mean (ComputeSaliencyMap(k))
20:
Compute nccBlank ← ComputeNCC (k, ${t e m p l a t e}_{b l a n k}$ )
21:
Compute nccLogo ← ComputeNCC (k, ${t e m p l a t e}_{l o g o}$ )
22:
if duration < ${Δ t}_{s h o r t}$ then
23:
Remove k from K
24:
else if nccBlank > Threshold_Blank or nccLogo > Threshold_Logo then
25:
Remove k from K
26:
else if personArea > AreaThreshold and not HasContext(k) then
27:
Remove k from K
28:
else if ssim > SSIMThresholdHigh and duration > DurationThreshold then
29:
Remove k from K
30:
else if ssim > SSIMThresholdModerate and saliencyScore < SaliencyThresholdLow then
31:
Add k to K’
32:
end if
33:
end for
34:
return K, K’