The Latest Technology of Pixelmator Pro: Enlarging Images Three Times Without Losing Quality

2019-12-19

In reality, we always encounter such dilemmas: some images are exquisite but have low resolution, and they become blurry when enlarged. Now, with the latest technology of Pixelmator Pro, enlarging images to three times the original resolution without any loss of clarity is no longer a problem.


Pixelmator Pro, a competitor of Photoshop, announced today its Pro version priced at $60, which features a function called ML Super Resolution. This function allows images to be scaled up to three times the original resolution without any image defects such as pixelation or blurring.


Here are some examples of Pixelmator:





How it works:


To create the ML Super Resolution function, engineers used convolutional neural networks. This type of deep neural network simplifies raster images and their complex pixel dependencies into something that is easy to process (i.e., requires less computation) without losing important features such as edges, patterns, colors, textures, and gradients. The ML Super Resolution network consists of 29 convolutional layers that scan the image and create a version with over 100 channels, which includes a range of recognized features. It is then upscaled, processed, and converted back to a raster image. Below is a simplified representation of the neural network.



First, the input image undergoes basic edge detection through a high-pass filter. Then, the first convolutional layer reduces the size of these features and merges the data. In the "descriptor fusion" block, the image is scanned to find all JPEG compression blocks and fuse them with other features identified so far.


Subsequent convolutional layers and residual blocks detect features (edges, patterns, colors, textures, gradients, etc.) in the image and build them into a complex representation with a depth of over 100 channels. In convolutional neural networks, more layers mean better accuracy, but if there are too many layers, the network becomes almost impossible to train. The residual blocks aim to increase the complexity and accuracy of the network while making it trainable.


Finally, all the features identified by the neural network are upscaled in the upscale block. Afterward, two residual blocks and the final convolutional layer perform post-processing on the data and then transform the features back into an image. It is also worth noting that all these operations occur on the device, and the entire trained machine learning model is included in the Pixelmator Pro application package.


Based on our tests, overall, the performance of Pixelmator's Super Resolution function is satisfactory. Various images, from illustrations to photos to text, can be smoothed out without pixelation. The results are better than what traditional upscaling algorithms provide. Although the process is not instantaneous (each image took about eight seconds on our 2017 MacBook Pro), the speed is sufficient to benefit designers and image editors in various industries.

RECENT AI NEWS

RECENT AI TOOLS