Image recognition with a wide area recognition target network that feeds a narrow area recognition target network
View Patent ↗Aspects relate to a processing device for increasing the accuracy of image recognition in a neural network that does not include a fully connected layer. A first processor generates a first feature map by executing processing of a first neural network on a target image. An enlarger enlarges the first feature map. A combiner combines the first feature map and the target image and generates a combined image. A second processor generates a second feature map by executing processing of a second neural network on the combined image.
1 . A processing device adapted to execute inference processing after learning processing, comprising:
a computer-readable non-transitory recording medium storing executable instructions that, in response to execution, causes the processing device to perform operations comprising:
processing, in the inference processing after learning, of a first neural network on a target image to be processed, thereby generating a first feature map having a smaller size than the target image,
up-sampling, in the inference processing after learning, the first feature map to have the same size as the target image,
concatenating, in the inference processing after learning, the first feature map that was up-sampled and the target image, thereby generating a combined image, and
processing, in the inference processing after learning, of a second neural network on the combined image, thereby generating a second feature map having a smaller size than the target image and a larger size than the first feature map,
wherein the first neural network does not include a fully connected layer, and the second neural network does not include a fully connected layer,
in the learning processing, first-stage learning is performed only on the first neural network, and
in the learning processing, a coefficient of a spatial filter derived by the first-stage learning is set to each convolution layer included in the first neural network, and
second-stage learning is performed on the second neural network after the first-stage learning has been performed.
2 . The processing device according to claim 1 , wherein the concatenating combines two inputs as different channels.
3 . A processing method performed by a processing device adapted to execute recognition processing after learning processing, the method including:
a computer-readable non-transitory recording medium storing executable instructions that, in response to execution, causes the processing device to perform the processing method, the processing method comprising:
a step of executing processing of a first neural network on a target image to be processed and generating a first feature map having a smaller size than the target image;
a step of up-sampling the generated first feature map to have the same size as the target image;
a step of concatenating the first feature map that was up-sampled and the target image to generate a combined image; and
a step of executing processing of a second neural network on the generated combined image and generating a second feature map having a smaller size than the target image and a larger size than the first feature map,
wherein the first neural network does not include a fully connected layer, and the second neural network does not include a fully connected layer,
in the learning processing, first-stage learning is performed only on the first neural network, and
in the learning processing, a coefficient of a spatial filter derived by the first-stage learning is set to each convolution layer included in the first neural network, and
second-stage learning is performed on the second neural network after the first-stage learning has been performed on the first neural network.