Deep-Learning Object Detection via Heat Map Segmentation
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Conventional image search systems fail to accurately identify specific objects within visual media due to cluttered backgrounds and distortions, and suffer from data scarcity issues that hinder the training of machine-learning algorithms for object recognition.
Innovation Solution
A deep-learning neural network architecture that generates heat maps to locate query objects within search media and expands the training dataset by creating synthetic images through augmentation, allowing for improved object detection and classification despite distortions and clutter.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If conventional image search is used to find a desired object in visual media, then the search process is simple, but the system cannot accurately identify the desired object when multiple objects are present or when the object is distorted
Solution Approach 1:
The patent segments the visual medium into multiple candidate regions and processes each region separately through the neural network. The system divides the image into grid cells, identifies candidate objects in each cell, and generates individual heat maps for each candidate before aggregating them. This segmentation allows the system to handle multiple objects and distortions effectively by focusing computational resources on relevant regions.
Solution Approach 2:
The patent introduces heat maps as an intermediary representation between the input image and the final object identification. The neural network generates heat maps that highlight the likelihood of object presence at each pixel location, serving as a mediator that transforms the complex image data into a simplified probability distribution that can be easily thresholded and processed to identify the desired object.
2Measurement precision
If the training dataset is expanded to include more varied images for better object recognition, then the recognition accuracy improves, but the data collection and processing time increases
Solution Approach 1:
The patent uses data augmentation techniques that create synthetic copies and transformations of existing training images. The system generates augmented versions of training images by applying transformations such as rotations, flips, crops, and color adjustments, effectively multiplying the training dataset without requiring additional manual data collection. This copying approach maintains recognition accuracy while minimizing time investment.
Solution Approach 2:
The patent implements data augmentation as a preliminary processing step before training the neural network. By pre-generating augmented versions of training images and preparing the expanded dataset in advance, the system eliminates the need for time-consuming data collection during the training process. The augmented data is ready for immediate use, significantly reducing the overall time required for model development.
Data Source
AI summary
Systems and methods are disclosed herein for automatically identifying a query object within a visual medium. The technique generally involves receiving as input to a neural network a query object and a visual medium including the query object. The technique also involves generating, by the neural network, representations of the query object and the visual medium defining features of the query object and the visual medium. The technique also involves generating, by the neural network, a heat map using the representations. The heat map identifies a location of pixels corresponding to the query object within the visual medium and is usable to generate an updated visual medium highlighting the query object.