Generic object-based image recognition apparatus with exclusive classifier, and method for the same
The present invention provides an image recognition apparatus with enhanced performance and robustness. In an image recognition apparatus 1 , an image classification information accumulating unit 20 stores therein feature information defining visual features of various objects obtained through a learning process. For classification of input images, an image feature obtaining unit 18 extracts descriptors indicating features from each input image, image vocabularies corresponding to the descriptors are voted, and a classifying unit 19 calculates existence probabilities of one or more candidate objects, based on the result of the voting. According to the existence probabilities, objects contained in the image is identified. Through the calculation, the existence probabilities are adjusted by an exclusive classifier, based on exclusive relationship information defining exclusive object sets each containing different objects (object labels) predicted not to coexist in a same image.
1. An image recognition apparatus for generic object recognition, comprising:
an identification unit identifying one or more objects contained in an input image by calculating existence probabilities of candidate objects by using LASSO (Least Absolute Shrinkage and Selection Operator) prediction with reference to feature information, the feature information defining visual features of various objects; and
an information accumulating unit accumulating therein exclusive relationship information, the exclusive relationship information defining one or more exclusive object sets each containing different objects predicted not to coexist in a same image, wherein
the identification unit adjusts the existence probabilities by using the exclusive relationship information, and
the adjustment is performed by setting a penalty value to a LASSO regularization parameter and thereby lowering existence probabilities of candidate objects that match any of the exclusive object sets defined by the exclusive relationship information.
2. The image recognition apparatus of claim 1 , wherein
the exclusive relationship information defines a plurality of exclusive object sets, and
the penalty value is set greater when a greater number of candidate object combinations match any of the exclusive object sets defined by the exclusive relationship information, each candidate object combination consisting of a predetermined number of candidate objects each having an existence probability no lower than a predetermined value.
3. The image recognition apparatus of claim 1 , wherein
the identification unit adjusts the existence probabilities by using the following formula:
(
w
^
,
u
^
)
-
arg
min
w
,
u
1
2
y
-
Xw
2
2
+
λ
2
∑
G
i
∈
G
u
G
i
1
2
,
where yεR m denotes a feature of the input image, XεR m×n denotes features of collected images as training images, n denotes a number of the collected images, m denotes a dimension number of descriptors, wεR n denotes a linear reconstruction coefficient for all object classes evaluated for the input image, G i denotes the exclusive label sets, and u Gi denotes the penalty value.
4. The image recognition apparatus of claim 1 , wherein
the exclusive relationship information is obtained by a statistical method from images as training images each containing identified objects.
5. The image recognition apparatus of claim 1 , wherein
the exclusive relationship is obtained by a graph shift method from images as training images each containing identified objects.
6. An image recognition method for generic object recognition, comprising:
an information accumulating step of accumulating exclusive relationship information, the exclusive relationship information defining one or more exclusive object sets each containing different objects predicted not to coexist in a same image; and
an identification step of identifying one or more objects contained in an input image by calculating existence probabilities of candidate objects by using LASSO (Least Absolute Shrinkage and Selection Operator) prediction with reference to feature information, the feature information defining visual features of various objects, wherein
through the calculation performed in the identification step, the existence probabilities are adjusted by using the exclusive relationship information, and
the adjustment is performed by setting a penalty value to a LASSO regularization parameter and thereby lowering existence probabilities of candidate objects that match any of the exclusive object sets defined by the exclusive relationship information.
7. A non-transitory computer-readable medium having a computer program stored theroen, the computer program for use with an image recognition apparatus for generic object recognition, and causing a computer to perform:
an information accumulating step of accumulating exclusive relationship information, the exclusive relationship information defining one or more exclusive object sets each containing different objects predicted not to coexist in a same image; and
an identification step of identifying one or more objects contained in an input image by calculating existence probabilities of candidate objects by using LASSO (Least Absolute Shrinkage and Selection Operator) prediction with reference to feature information, the feature information defining visual features of various objects, wherein
through the calculation performed in the identification step, the existence probabilities are adjusted by using the exclusive relationship information, and
the adjustment is performed by setting a penalty value to a LASSO regularization parameter and thereby lowering existence probabilities of candidate objects that match any of the exclusive object sets defined by the exclusive relationship information.
8. An integrated circuit used in an image recognition apparatus for generic object recognition, comprising:
an identification unit identifying one or more objects contained in an input image by calculating existence probabilities of candidate objects by using LASSO (Least Absolute Shrinkage and Selection Operator) prediction with reference to feature information, the feature information defining visual features of various objects; and
an information accumulating unit accumulating therein exclusive relationship information, the exclusive relationship information defining one or more exclusive object sets each containing different objects predicted not to coexist in a same image, wherein
the identification unit adjusts the existence probabilities by using the exclusive relationship information, and
the adjustment is performed by setting a penalty value to a LASSO regularization parameter and thereby lowering existence probabilities of candidate objects that match any of the exclusive object sets defined by the exclusive relationship information.