Convolution atrous
La convolution atrous, également connue sous le nom de convolution dilatée, is a technique used in apprentissage profond, particularly in réseaux de neurones convolutifs (CNNs). It allows for the expansion of the receptive field without losing resolution or increasing the number of parameters de manière significative.
In standard convolution, a filter or kernel slides over the input data, computing a produit scalaire at each position. In atrous convolution, the filter is applied with gaps (or ‘holes’) between its weights. This means that instead of covering every input pixel, the filter skips some, effectively ‘dilating’ its size. By adjusting the spacing between the filter weights, atrous convolution can capture features at multiple scales while maintaining the same output size as the original input.
This method is particularly useful in tasks such as image segmentation, where understanding context from various resolutions is important. For example, in segmentation sémantique, atrous convolution allows the model to consider larger regions of context without downsizing the input image or losing information about finer details.
Atrous convolution is often used in conjunction with other techniques, such as skip connections and residual networks, to improve feature extraction and performance globale of deep learning models. The dilation rate, which determines the spacing of the filter, can be adjusted to control how much context the model considers, making it a flexible and powerful tool in modern neural network architectures.