Neural Network Models for Semantic Image Segmentation

A computer processing system for training a model for use in semantic image segmentation,

• wherein the model comprises:
• a first classifier neural network;
• and
• a second classifier neural network,
• wherein the computer processing system is configured to train the first classifier neural network by:
• inputting a first training image containing a target object to the first classifier neural network;
• using the first classifier neural network to identify pixels of the first training image that are discriminative for the target object;
• generating a first modified training image by attenuating or erasing the identified pixels of the first training image;
• inputting the first modified training image to the second classifier neural network;
• using the second classifier neural network to determine a likelihood of the first modified training image containing the target object;
• and
• updating weights of the first classifier neural network using a first loss function that is a monotonically-increasing function of the determined likelihood, and wherein the computer processing system is configured to train the second classifier neural network by:
• inputting a second training image, containing the target object, to the first classifier neural network;
• using the first classifier neural network to identify pixels of the second training image that are discriminative for the target object;
• generating a second modified training image by attenuating or erasing the identified pixels of the second training image;
• inputting a set of training images to the second classifier neural network, the set comprising the second modified training image and one or more training images that do not contain the target object;
• and
• for each training image of the set, using the second classifier neural network to determine a likelihood of the respective training image containing the target object, and updating weights of the second classifier neural network using a second loss function, different from the first loss function, wherein the second loss function is a monotonically-decreasing function of the determined likelihood when the training image contains the target object, and is a monotonically-increasing function of the determined likelihood when the training image does not contain the target object.

Computer software comprising instructions which, when executed on a computer processing system, cause the computer processing system to train a model for use in semantic image segmentation, wherein the model comprises:

• a first classifier neural network;
• and
• a second classifier neural network,
• wherein the instructions cause the computer processing system to train the first classifier neural network by:
• inputting a first training image containing a target object to the first classifier neural network;
• using the first classifier neural network to identify pixels of the first training image that are discriminative for the target object;
• generating a first modified training image by attenuating or erasing the identified pixels of the first training image;
• inputting the first modified training image to the second classifier neural network;
• using the second classifier neural network to determine a likelihood of the first modified training image containing the target object;
• and
• updating weights of the first classifier neural network using a first loss function that is a monotonically-increasing function of the determined likelihood, and
• wherein the instructions cause the computer processing system to train the second classifier neural network by:
• inputting a second training image, containing the target object, to the first classifier neural network;
• using the first classifier neural network to identify pixels of the second training image that are discriminative for the target object;
• generating a second modified training image by attenuating or erasing the identified pixels of the second training image;
• inputting a set of training images to the second classifier neural network, the set comprising the second modified training image and one or more training images that do not contain the target object;
• and
• for each training image of the set, using the second classifier neural network to determine a likelihood of the respective training image containing the target object, and updating weights of the second classifier neural network using a second loss function, different from the first loss function, wherein the second loss function is a monotonically-decreasing function of the determined likelihood when the training image contains the target object, and is a monotonically-increasing function of the determined likelihood when the training image does not contain the target object.