Kaggle Whale Challenge jangcs nthu edu tw http

Kaggle: Whale Challenge 張智星 jang@cs. nthu. edu. tw http: //www. cs. nthu. edu. tw/~jang 多媒體資訊檢索實驗室 台灣大學 資訊 程系

Whale Challenge z Problem definition y. Identify the existence of whales from sensor recordings z Characteristics: y. Imbalance data y. Some recordings are hardly recognizable by non-experts -2 -

Dataset z. Training set y 47, 844 recordings of 2 seconds x 88. 97% (42, 565 recordings): w/o whales x 11. 03% (5, 276 recordings): with whales z. Test set y 25, 468 recordings of 2 seconds z. Recording format y 2000 -Hz sample rate, 16 -bit resolution -3 -

Preprocessing z. Potential preprocessing y. Trend removal x. Trend estimation via polynomial fitting y. Noise removal x. Band-pass filter y. Removal of “non-whale” part x. Linear prediction? -4 -

Spectrogram kwc. Preprocess. m z W/o band-pass filter z W/ band-pass filter -5 -

Potential Features z Acoustic features y. Volume y. Pitch y. Spectrum y. MFCC y… z Visual features (obtained from spectrogram) y. Radon transform y. Hough transform y. Gabor filters y… -6 -

Pitch Tracking z kwc. Pitch. Tracking. m -7 -

Volume z kwc. Volume. m -8 -

Spectrogram z kwc. Spectrogram. m -9 -

Visual Features via Radon Transform z Radon transform y. Projection onto lines at various angles y. For grayscale images only y. Detection objects at a specific angle -10 -

Example of Radon Transform z Source z Output Code: go. Radon. m yhttp: //www. mathworks. com/help/images/ref /radon. html -11 -

Example of Radon Transform (2) z Source image z Output Code: go. Radon 2. m -12 -

Visual Features via Hough Transform z Hough transform y. Commonly used for detection lines and circles y. For BW images only (after edge detection) -13 -

Visual Features via Hough Transform (2) z. Hough transform y. Point to curve mapping y. Two points Two sine curves x. The intersection is the right θ and ρ for the line connecting these two points -14 -

Example of Hough Transform z Source yhttp: //www. ebsdimage. org/documentation/reference/ops/hough/op/ho ughtransform. html Image Hough space and its maxima Detected lines -15 -

Example of Hough Transform (2) z Source yhttp: //www. mathworks. com/help/images/analyzingimages. html (MATLAB code available) Image Edge image Hough space and its maxima Detected lines -16 -

Methods z Thresholding y. Volume variance y. Pitch variance z Static classifiers z Sequence classifiers y. HMM y. CRF y… y. Naïve Bayes classifiers y. GMM y. SVM y… -17 -

HMM Training z kwc. Hmm. Train. m -18 -

HMM Evaluation z kwc. Hmm. Eval. m -19 -

HMM z Basic models z Advanced models y. Class 1: sil ysil-whale-sil-whalesil y… 1. 0 sil y. Class 2: sil-whale-sil 0. 9 sil 0. 4 0. 1 w 1. 0 0. 6 sil -20 -

HMM (2) z. Other approach y. Train HMM models y. Align each recording with the HMM 0. 9 sil 0. 4 0. 1 w 1. 0 0. 6 sil y. Extract features from the whale part for other static classifiers x. Duration (no. of frames) x. Average log likelihood per frame -21 -

Performance Evaluation z. Performance evaluation of methods based on thresholding (http: //en. wikipedia. org/wiki/Receive r_operating_characteristic): y. ROC, DET y. AUC -22 -