Back to results
US 8,897,575 B2Granted
Multi-scale, perspective context, and cascade features for object detection
AssigneeSeiko Epson Corporation
InventorsYuanyuan Ding, Jing Xiao
Life of this patent
Estimating the term…
Abstract
Systems and methods for object detection are presented herein. Embodiments of the present invention utilizing a cascade feature, one or more features at different scales, one or more multi-scale features in combination with a perspective feature, or combinations thereof to detect an object of interest in an input image. In embodiments, the various features are used to train classifiers. In embodiments, the trained classifiers are used in detecting an object of interest in one or more input images.
Claims15
A non-transitory computer-readable medium or media comprising one or more sequences of instructions which, when executed by one or more processors, causes steps to be performed comprising:
- for each image patch from a set of one or more image patches from an image:
- generating a context region for the image patch comprising at least a portion of the image;
- and
- constructing a multi-scale feature comprising two or more of:
- a high-scale feature comprises one or more features extracted from the image patch at a high-resolution size;
- a mid-scale feature comprises one or more features extracted from the image patch at a mid-resolution size;
- and
- a low-scale feature comprises one or more features extracted from a combined image patch at a low-resolution size, the combined image patch comprising the image patch and the context region;
- wherein the mid-scale feature comprises a combination of the one or more features extracted from the image patch at the mid-resolution size and one or more features extracted from the context region;
- and
- wherein the one or more features extracted from the context region comprises:
- forming the context region into a plurality of regions;
- for each region from the plurality of regions that does not have a mid-resolution size, re-sizing the region to the mid-resolution size;
- extracting one or more features for each region of the plurality regions;
- and
- combining the one or more features from each of the regions to form the one or more features extracted from the context region.
13 more elements
A processor-based system for detecting an object in an image, the system comprising:
- an image patch generator for receiving an input image and for generating an image patch from the input image that may contain the object and a context region for the image patch;
- a feature generator, communicatively coupled to receive the image patch, for generating a multi-scale feature comprising at least two of:
- a high-scale feature comprising one or more features extracted from the image patch at a high-resolution size;
- a mid-scale feature comprising a combination of one or more features extracted from the image patch at a mid-resolution size and one or more features extracted from a set of one or more regions at the mid-resolution size, the set of one or more regions being formed from the context region;
- and
- a low-scale feature comprises one or more features extracted from a combined image patch at a low-resolution size, the combined image patch comprising the image patch and the context region;
- and
- a detector, communicatively coupled to receive the multi-scale feature for the image patch, for using a trained classifier corresponding to multi-scale feature to provide a score related to whether to image patch contains the object;
- wherein the generating of the combination of one or more features extracted from the image patch at a mid-resolution size includes:
- dividing the context region into a plurality of regions;
- for each region from the plurality of regions that does not have the mid-resolution size, re-sizing the region to the mid-resolution size;
- extracting one or more features for each region of the plurality of regions;
- and
- combining the one or more features from each of the regions to form the combination of one or more features extracted from the context region.
11 more elements
A non-transitory computer-readable medium or media comprising one or more sequences of instructions which, when executed by one or more processors, causes steps to be performed comprising:
- for each image window from a set of one or more image windows from the image, defining a context region for the image window by selecting a portion or portions of the image that are within a context region area defined relative to the image window;
- and
- for each image window from the set of one or more image windows, constructing a multi-scale feature descriptor comprising a combination of a high-scale feature, a middle-scale feature, and a low-scale feature, wherein:
- the high-scale feature comprises one or more features extracted from the image window at a first resolution;
- the middle-scale feature comprises a combination of one or more features extracted from the image window at a second resolution and one or more features extracted from a set of one or more regions at the second resolution, the set of one or more regions being formed from the context region;
- and
- the low-scale feature comprises one or more features extracted from a combined image window having a third resolution, the combined image window comprising the image window and the context region;
- wherein the step of forming the combination of one or more features extracted from a set of one or more regions at the second resolution comprises:
- dividing the context region into a plurality of regions;
- for each region from the plurality of regions that does not have the second resolution size, re-sizing region to the re the second resolution size;
- extracting one or more features for each region of the plurality of regions;
- and
- combining the one or more features from each of the regions to form the combination of one or more features extracted from the context region.
10 more elements
Dependencies are inferred from the claim text.