Learning Apparatus for Object Density Measurement via Feature Extraction
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Conventional techniques for calculating object density from images face challenges with high memory requirements due to large data volumes and reduced detection accuracy for small objects or those with overlap in images.
Innovation Solution
A learning apparatus that extracts partial images, calculates feature amounts, and learns a regression model using a random forest to predict object density with low memory capacity by reducing variation in voting histograms and predicting representative labels.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If conventional random forest learning is used to calculate object density, then density measurement capability is achieved, but memory capacity requirements become enormous due to large data volume
Solution Approach 1:
The patent extracts only the essential features needed for density calculation by using HOG (Histogram of Oriented Gradients) to capture shape and texture information, and by representing objects with bounding box coordinates rather than storing complete image data. This extraction approach maintains measurement capability while dramatically reducing memory requirements.
Solution Approach 2:
Instead of storing large amounts of raw image data and processing it during density calculation, the patent inverts the approach by pre-processing images to extract compact feature representations (HOG descriptors and bounding boxes) that can be efficiently stored and quickly queried during measurement.
2Ease of operation
If conventional object detection techniques are used, then detection capability is achieved, but detection accuracy decreases when objects are small or have large overlap in images
Solution Approach 1:
The patent segments the detection process into multiple stages: first detecting candidate regions, then refining detections through non-maximum suppression, and finally calculating density from confirmed object positions. This segmented approach improves accuracy for small and overlapping objects by allowing progressive refinement of detection results.
Solution Approach 2:
The patent implements feedback mechanisms through iterative detection refinement and non-maximum suppression, where detection results are continuously improved by comparing candidate detections against learned patterns and eliminating redundant or low-confidence detections based on overlap analysis.
Data Source
AI summary
A first extracting unit extracts partial images from a learning image. A first calculator calculates a feature amount of the partial image. A retrieving unit retrieves objects included in the partial image, and gives, as a label, a vector to the feature amount. The vector represents relative positions between a first position in the partial image and each object included in the partial image. A voting unit generates a voting histogram for each partial image. A learning unit divides the feature amount of each partial image into clusters to reduce variation of the corresponding voting histogram, so as to learn a regression model representing a relationship between the feature amount of the partial image and the relative position of the object included in the partial image. A first predicting unit predicts, for each cluster, a representative label from the label given to the feature amount belonging to the cluster.