ECE 539 Project Report Birdspecies Recognition Using Convolutional

  • Slides: 11
Download presentation
ECE 539 – Project Report Bird-species Recognition Using Convolutional Neural Network -- Instructor: Yu

ECE 539 – Project Report Bird-species Recognition Using Convolutional Neural Network -- Instructor: Yu Hen Hu -- Presenter: Xucheng Wan

Problem Introduction Computer Vision → Recognition & Classification → Datasets

Problem Introduction Computer Vision → Recognition & Classification → Datasets

Caltech UCSD Bird 200 - 2011 (CUB-200) (Established for subordinate categorization study) 200 Bird-species

Caltech UCSD Bird 200 - 2011 (CUB-200) (Established for subordinate categorization study) 200 Bird-species * 60 images for each

Method and Progress I. Bird Recognition - Binary Classifier

Method and Progress I. Bird Recognition - Binary Classifier

I. Binary Classifier Data: Cifar(60 k)+ CUB-200(12 k) Validation: 20% Accuracy: 0. 951

I. Binary Classifier Data: Cifar(60 k)+ CUB-200(12 k) Validation: 20% Accuracy: 0. 951

II. Multi-Classifier # Classes increases → need more features → # Layers increase

II. Multi-Classifier # Classes increases → need more features → # Layers increase

II. Multi-Classifier # of Classes Training Accuracy Validation Accuracy 20 0. 87 0. 637

II. Multi-Classifier # of Classes Training Accuracy Validation Accuracy 20 0. 87 0. 637 80 0. 68 0. 397 200 0. 42 0. 273 Data is so limited !!!! Only 60 images for each of 200 bird-species.

III. Revised Goog. Le. Net Modified some layers and activation functions

III. Revised Goog. Le. Net Modified some layers and activation functions

III. Revised Goog. Le. Net Dataset Subset 1 Class 1 ··· Class k Subset

III. Revised Goog. Le. Net Dataset Subset 1 Class 1 ··· Class k Subset 10 Group Name Validation Loss Validation Accuracy Entire Dataset 0. 5718 0. 8202 Weighted Average 0. 4664 0. 8697

Conclusion & Evaluation Method Validation Accuracy Bird Recognition with Binary Classifier 0. 951 CNN

Conclusion & Evaluation Method Validation Accuracy Bird Recognition with Binary Classifier 0. 951 CNN with 5 Layers and Soft. Max 0. 2730 Revised Goog. Le. Net and Soft. Max 0. 4771 Revised Goog. Le. Net on 10 subsets 0. 7133 1. Multi-output activation > Binary classifier 2. More layers VS Overfitting 3. Data is the most important

Thank You ! Questions ?

Thank You ! Questions ?