Neural Networks. 2.4.1 Convolutional Neural Network
Neural Networks. In this section, we will introduce the neural networks knowledge such as word embeddings, convolutional neural networks, long short-term memory networks, and attention mechanism. We will use these techniques in the later chapter to develop our novel model.
Neural Networks. Neural networks, though not an area of significant active research at present, were included within this project as they represent an accurate, based upon the literature review, and interesting method for visual object matching, allowing the modelling of nonlinear multivariant information. Neural networks also have the advantage of being able to create object class invariants of sorts. If objects of the same class, but at different rotations (though equally applicable to lighting variations and scales, etc.), are presented to the network during its training phase, all differences between the class objects will be assimilated (with an associated error rating) into the model used, and objects at these different rotations will all be treated as the same class. This naturally requires painstaking effort during the creation of the training set and often requires bootstrapping of the training set (i.e. repeatedly finding and removing contentious images) to resolve an adequate network error rate. The main drawback to the use of neural networks, which became particularly apparent when implemented within this project, is the time required per classification when using large networks. There appears to be a “catch 22” type situation attached to the training of neural networks. On the one hand, the addition of training examples improves the classification accuracy of the network as a whole; however, the addition of new training examples slows the network during classification and increases the overall network training error making it more difficult and time consuming to train. An additional drawback of using neural networks is that the weighting file (trained during the original training process), cannot be easily updated without training the entire system for every object class again. This means that new objects (objects not included in the original specification for the project) cannot be simply added to a neural network based classification system without complete retraining. [85] assesses the suitability of possible image rescaling and normalisation methods, using the neural network technique, due to its sensitivity to these factors (the differences in rescaling and normalisation techniques created a 3% difference in results). Though useful for deciding upon preprocessing techniques, this type of sensitivity is not advantageous for object classification within a project such as AVITRACK. This sensitivity to image conditions stretches to include, and become pe...
