DeConvolution and How it works and its importance?
Before Deconvolution, let's quickly review What Convolution is.
Convolution is extracting important information from images using Kernel, and make an output matrix often known as a Feature map. Image’s dimensions get reduced and important information is also retained.
DeConvolution is totally opposite process from Convolution. It is also known as Transposed Convolution. In DeConvolution, the Feature map gets converted into an Image. Converting Convoluted image into Original Image is DeConvolution.
Convolution adds up the information spread in various pixels to one pixel, whereas DeConvolution spreads the information present in one pixel to various pixels.
This forcefully compels us to think that from where the extra pixels are coming when we convert the Convoluted image to the Original image. Other than that, It is not possible to insert one pixel too many.
The solution for this could be adding some Padding before applying Convolution, the output matrix will obviously look different without Padding. Padding will add lots of 0s at the border which could look weird in starting. But adding Padding helps us bring back the image of its Original Dimension. That is DeConvolution and it works.
Where to use DeConvolution?
Deconvolution is used for Image Segmentation. Image Segmentation is dividing an image into multiple segments or classes. Segmentation makes it easier to understand and analyze the images. Segmentation is a computationally very expensive process because we need to classify each pixel for this.
Since Segmentation is about finding classes of every pixel, here downsampled Feature map cannot work. So we use DeConvolution to convert it into an image of the same dimensions as of the Original Image.
DeConvolution is done after applying the Convolution layer, to maintain the Output size of the image same as Original Image.
Why DeConvolution is so important?
Segmentation is one of the uses of Deconvolution. In any Computer Vision application, where the resolution of the final image required is larger than the input, Deconvolution becomes important.
The DeConvolution layer is very popular in applications like Generative Adversarial Network(GANs), Image super-resolution, Surface depth estimation, etc. In such use cases, the DeConvolution layer is used to merge responses from different Convolutional layers and significantly boost the accuracy.
Stay tuned and Keep Learning.
Follow to learn many more interesting things on Instagram.
