Spaces:
Sleeping
Sleeping
Niharmahesh
commited on
Commit
•
769e0dd
1
Parent(s):
fa7427c
Create readme.md
Browse files
readme.md
ADDED
@@ -0,0 +1,36 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
# ASL Recognition App: Image Preprocessing and Feature Extraction
|
2 |
+
|
3 |
+
This document explains the image preprocessing and feature extraction techniques used in our ASL Recognition App.
|
4 |
+
|
5 |
+
## Image Preprocessing
|
6 |
+
|
7 |
+
1. **Image Loading**: We use OpenCV (cv2) to load and process images.
|
8 |
+
2. **Color Conversion**: Images are converted from BGR to RGB color space for compatibility with MediaPipe.
|
9 |
+
|
10 |
+
## Hand Landmark Detection
|
11 |
+
|
12 |
+
We use MediaPipe Hands for detecting hand landmarks in the images:
|
13 |
+
|
14 |
+
1. **MediaPipe Hands**: Initializes with `static_image_mode=True`, `max_num_hands=1`, and `min_detection_confidence=0.5`.
|
15 |
+
2. **Landmark Extraction**: 21 hand landmarks are extracted from each image.
|
16 |
+
|
17 |
+
## Feature Extraction
|
18 |
+
|
19 |
+
### Landmark Normalization
|
20 |
+
|
21 |
+
1. **Centering**: Landmarks are centered by subtracting the mean position.
|
22 |
+
2. **Scaling**: Centered landmarks are scaled to unit variance.
|
23 |
+
|
24 |
+
### Angle Calculation
|
25 |
+
|
26 |
+
We calculate angles between all pairs of landmarks:
|
27 |
+
|
28 |
+
1. **Vector Calculation**: For each pair of landmarks (i, j), we calculate the vector from i to j.
|
29 |
+
2. **Angle Computation**: We compute the arcosine of the x and y components of the normalized vector.
|
30 |
+
3. **Feature Vector**: The angles form a 420-dimensional feature vector (21 choose 2 = 210 pairs, 2 angles per pair).
|
31 |
+
|
32 |
+
## Model Input
|
33 |
+
|
34 |
+
The preprocessed features (angles) are used as input to our Random Forest model for ASL sign classification.
|
35 |
+
|
36 |
+
This preprocessing pipeline ensures that our model receives consistent and informative features, regardless of the hand's position or size in the original image.
|