Method for object detection using hierarchical deep learning

Inventors

Cagan, Jonathan • Leduc, Philip • Clymer, Daniel

Abstract

A hierarchical deep-learning object detection framework provides a method for identifying objects of interest in high-resolution, high pixel count images, wherein the objects of interest comprise a relatively a small pixel count when compared to the overall image. The method uses first deep-learning model to analyze the high pixel count images, in whole or as a patchwork, at a lower resolution to identify objects, and a second deep-learning model to analyze the objects at a higher resolution to classify the objects.

Core Innovation

The invention provides a hierarchical deep-learning object detection and classification approach for identifying one or more objects of interest in an image while analyzing the image at a lower resolution than the native resolution. A trained machine learning model identifies the objects of interest, and the identified objects of interest are then classified at a higher resolution using a trained machine learning classifier. The machine learning classifier outputs a latent vector for each classified object.

The latent vectors for a predetermined number of classified objects are aggregated, and the aggregated latent vectors are pooled by calculating a maximum or minimum of the data for each node of a feature map of the machine learning classifier. A binary classification is then performed for each image based on the aggregated latent vectors, and the binary classification indicates the presence or absence of an object of interest. This supports performing image-level classification using pooled aggregated latent representations.

In the described framework, the object identification at lower resolution provides inputs for higher-resolution object classification, and latent vectors from classified objects are combined through concatenation and pooled feature-map operations using per-node maximum or minimum. The invention further supports reducing the latent vector dimensionality by principal component analysis of pooled aggregated latent vectors and optionally combining the pooled representations with metadata for training an image-level binary classifier. The document describes achieving binary disease/healthy labeling for each image based on aggregated latent vectors.

Claims Coverage

The independent claim recites a hierarchical method that identifies objects at a lower resolution, classifies the identified objects at a higher resolution while producing latent vectors, aggregates and pools those latent vectors by per-node maximum or minimum on a feature-map, and performs an image-level binary classification indicating presence or absence of an object of interest. Dependent claims refine the approach by specifying additional model components and operations including dimensionality reduction, patch-wise analysis, and alternative classifier implementation.

Overall, the claim set centers on hierarchical lower-resolution object identification followed by higher-resolution object classification that produces latent vectors, and an image-level binary decision produced from predetermined-number latent vector aggregation with per-node maximum or minimum pooling. Dependent claims further narrow the implementation by specifying patch-wise analysis, principal component analysis on pooled aggregated latent vectors, and performing the binary classification using a random decision forest, among other refinements explicitly described in the provided claim excerpts.

Stated Advantages

Documented Applications