Image recognition apparatus

1. An image recognition apparatus configured to extract a target in time-series frame images through image processing for the frame images and output target information related to the target, the frame images being obtained by photographing an ambience from a moving object, the image recognition apparatus comprising a controller configured to:

perform positional detection and identification for the target in each of the frame images; and

extract a first target having an ambience change feature with priority over a second target that does not have the ambience change feature, the ambience change feature being a feature about a positional change of the first target that is exhibited when the ambience is photographed from the moving object, the positional change being a positional change of the first target identified in common among the time-series frame images;

wherein:

the controller includes a neural network programmed to recognize the target by using information on the time-series frame images; and

the neural network is programmed to learn through deep learning such that the first target having the ambience change feature is extracted with priority over the second target that does not have the ambience change feature; and

wherein, when the neural network learns through the deep learning, an error to be output from a loss function for updating a synaptic weight in the neural network is adjusted to be smaller in the first target having the ambience change feature than the second target that does not have the ambience change feature.

2. The image recognition apparatus according to claim 1 , wherein the ambience change feature includes a feature of a change in a size of the first target in addition to the feature of the positional change of the target.

3. The image recognition apparatus according to claim 1 , wherein the ambience change feature includes a feature that a positional change between the frame images is equal to or smaller than a predetermined amount for a first target recognized at a position near a vanishing point of a motion vector.

4. The image recognition apparatus according to claim 1 , wherein the ambience change feature includes a feature that the first target identified in common moves along a straight line.

5. The image recognition apparatus according to claim 4 , wherein the ambience change feature includes a feature that the target identified in common among the frame images moves along a straight line connecting the first target in an arbitrary frame image and a vanishing point of a motion vector.

6. The image recognition apparatus according to claim 4 , wherein the ambience change feature includes a feature that a change in a movement interval of the first target identified in common has a predetermined regularity.

7. The image recognition apparatus according to claim 4 , wherein the ambience change feature includes a feature that a change in a size of the first target identified in common has a predetermined regularity.

8. The image recognition apparatus according to claim 1 , wherein the controller is configured to:

input the time-series frame images obtained by photographing the ambience from the moving object;

set, as a target for the image processing, a frame image extracted at a predetermined sampling interval from the input frame images; and

adjust the sampling interval such that the sampling interval increases as a moving speed of the moving object decreases.

9. The image recognition apparatus according to claim 1 , wherein the moving object is a vehicle.

10. An image recognition apparatus comprising a controller configured to:

perform positional detection and identification for a target in each of time-series frame images obtained by photographing an ambience from a moving object;

extract the target in the frame images through image processing for the frame images, the controller being configured to extract a first target having an ambience change feature with priority over a second target, the ambience change feature being a feature about a positional change of the target that is needed for the target identified by the controller to be estimated as a common target among the time-series frame images when the ambience is photographed from the moving object; and

output target information related to the first target or the second target;

wherein:

the controller includes a neural network configured to recognize the target by using information on the time-series frame images; and

the neural network is configured to learn through deep learning such that the first target having the ambience change feature is extracted with priority over the second target that does not have the ambience change feature; and

wherein, when the neural network learns through the deep learning, an error to be output from a loss function for updating a synaptic weight in the neural network is adjusted to be smaller in the first target having the ambience change feature than the second target that does not have the ambience change feature.

11. The image recognition apparatus according to claim 10 , wherein the moving object is a vehicle.