Bing has announced version 3 of its image search engine’s “aesthetics” model, which brings more aesthetically pleasing results because it takes into account criteria such as image size, but not contrast, resolution or size. ‘Exposure of the file …
Bing recently announced improvements to its image search engine with version 3 of its “Aesthetic Model.” Objective . to use the model of read or learn attentively Deliver stunning and aesthetically pleasing photos using pre-trained encoders, web-scale data, and multitasking training goals.
V2 model improved Deep Attraction Rank Net (DARN) To directly find the mean and variance of the attractiveness scores of each image (DARN-V2) with human marking (crowdsourcing). In July of this year, the V3 improved upon some of the limitations of the previous model. It is particularly sensitive to the size of the image. In Model V2, the images were resized to a uniform size before training, causing distortion in the images in the training dataset. The V3 model is sensitive to size, as size is an essential feature in determining the aesthetics of an image. The original shape signal allows the model to compensate for the degradation and image distortion experienced by the V2 model. Furthermore, the model gives a higher score for a larger image if the images are similar but different sizes. The V3 model also takes into account the file’s composition, subjects, brightness, contrast, exposure, focus, saturation, and resolution.
You can read more about it on the Bing blog (but be careful, it’s quite complicated).
Comparison between engine results in V2 and V3. Source: Bing.
Prone to fits of apathy. Music specialist. Extreme food enthusiast. Amateur problem solver.