Neural Network Object Classification via Multiple Instance Learning
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Deep learning models face challenges in accurately detecting and classifying small objects within images due to the lack of patch-level labels and computational inefficiencies when training on whole slide images, which are computationally expensive and require costly annotations.
Innovation Solution
The implementation of a multiple instance learning (MIL) approach using convolutional neural networks (CNNs) with self-attention transformer blocks to infer instance features and pseudo-labels, allowing for patch-level training with bag-level labels, and utilizing a combination of bag-level and patch-level losses to improve classification accuracy.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If deep learning models are trained on whole slide images to detect small objects, then detection capability is improved, but computational cost and annotation expense increase significantly
Solution Approach 1:
The patent divides whole slide images into smaller image patches, allowing the model to process local regions independently. This segmentation reduces the computational burden of processing entire high-resolution images while maintaining detection accuracy for small objects through focused local analysis.
Solution Approach 2:
The patent introduces a multiple instance learning framework that operates at both the patch level and image level, adding a hierarchical dimension to the processing architecture. This allows the model to learn from multiple patches and aggregate their predictions, reducing the need to process every pixel at full resolution.
2Productivity
If patch-level labels are used for training, then training efficiency is improved, but label accuracy and reliability deteriorate due to lack of fine-grained annotations
Solution Approach 1:
The patent uses image-level labels as intermediary supervision signals to guide the learning process at the patch level. The multiple instance learning framework aggregates patch-level predictions and compares them against reliable image-level annotations, allowing the model to learn from coarse-grained labels while maintaining training efficiency.
Solution Approach 2:
The model performs self-supervised learning by generating its own pseudo-labels through the multiple instance learning process. The aggregation of patch-level predictions provides self-verified supervision signals that improve label reliability without requiring expensive fine-grained annotations.
3Area of stationary object
If small objects occupy only a very small area in images, then image resolution is maintained, but detection difficulty increases due to limited object presence
Solution Approach 1:
By segmenting the image into multiple patches, the model increases the probability that small objects will be captured in at least some patches. This allows the model to focus computational resources on patches containing objects of interest, improving detection difficulty despite limited object coverage.
Solution Approach 2:
The patent applies different processing strategies to different patches based on their content. Patches containing small objects receive focused attention through the multiple instance learning mechanism, while patches without objects are processed more efficiently. This local quality differentiation improves detection of small objects without compromising overall image coverage.
Data Source
AI summary
Apparatuses, systems, and techniques are presented to classify objects in images. In at least one embodiment, one or more neural networks are used to identify one or more objects in one or more full images based, at least in part, on the one or more neural networks having been trained using the one or more full images and one or more portions of the one or more full images.