G06V 10/82

Definition

(Für diese Definition ist die deutsche Übersetzung noch nicht abgeschlossen)

Neural networks (NN) specially adapted for image or video recognition or understanding, in particular specific architectures and specific learning tasks for this purpose.

Notes – technical background

These notes provide more information about the technical subject matter that is classified in this place:

Examples of architectures are:

• Attention based neural networks such as transformer architectures;
• Autoencoders consisting of encoder and decoder blocks, where the output has the same form as the input, i.e. input and outputs are both images for example;
• Convolutional neural networks consisting of repetitive convolutional and pooling layers;
• Pyramidal or multi-scale neural network, mostly of the convolutional type that process either differently scaled input images, have convolutional kernels of varying sizes, and/or contain skip connections from lower level layers to higher level layers or the output layer;
• Recurrent neural networks, where the input data is sequential by nature. That is either the pixels of the input image are processed sequentially, or a plurality of image frames such as in videos are processed sequentially. Long-short-term-memory (LSTM) and gated recurrent units (GRU) are some specific examples of recurrent neural networks;
• Region proposal networks, where the main task is not only to correctly classify objects in an input image but also provide an indication where a specific object has been found. Example architectures are R-CNN and YOLO;
• Residual neural networks (ResNet) containing skip connections or shortcuts to jump over some layers;
• Siamese neural networks, that work on input pairs and consist of two identical neural networks that process each pair of the input and then merges the output to provide a judgement about the input pair, such as if they are belonging to the same class or not.

Examples of learning tasks are:

• Adversarial learning such as in generative adversarial networks (GANs);
• Meta learning;
• Metric learning, learning a distant metric between two input objects mostly done with a Siamese neural network;
• Reinforcement learning, learning how to take optimal actions for performing a task, e.g. deep reinforcement learning for robotics, self-driving vehicles etc.;
• Representation or feature learning, learning representations or features from raw input, mostly done with some form of encoder-decoder architecture or simply by using intermediate representations of a classification network;
• Transfer or multitask learning, reusing a network trained on task A for task B or jointly training a neural network on multiple tasks.

Examples

Siamese network showing not similar (left) and similar (right) input pairs

Recurrent neural network for action recognition

Region proposal neural network for region of interest (ROI) detection

Adversarial learning with a generative adversarial neural network for object recognition on different backgrounds

Querverweise

Informative Querverweise

Glossar