Learning Multi-Instance Deep Discriminative Patterns for Image Classification.

Area of Science:

• Computer Vision
• Machine Learning
• Artificial Intelligence

Background:

• Effective image representation is crucial for image classification.
• Traditional methods like bag-of-features aggregate local descriptors, but weakly supervised part-based models offer improved discriminative power by using image labels.

Purpose of the Study:

• To propose a novel weakly supervised strategy for learning discriminative image representations using multi-instance learning (MIL).
• To extend traditional MIL methods to explicitly learn multiple patterns within positive classes and identify the most representative instance for each pattern.

Main Methods:

• A weakly supervised strategy employing multi-instance learning (MIL) is proposed.
• The method extends traditional MIL by learning multiple patterns in positive classes and identifying the 'most positive' instance for each.
• Local descriptors from deep convolutional neural networks (CNNs) are utilized for enhanced discriminative power, with positiveness treated as a continuous variable optimized via stochastic gradient descent.

Main Results:

• The proposed method demonstrates remarkable performance on widely used image classification benchmarks, including Action 40, Caltech 101, Scene 15, MIT-indoor, and SUN 397.
• The use of deep convolutional neural network features significantly improves the discriminative capability of the learned patterns.

Conclusions:

• The developed weakly supervised MIL strategy effectively learns discriminative patterns for image representation.
• The approach offers a significant advancement in image classification by leveraging deep learning features and enhanced MIL techniques.