Pneuma.to

光線穿過鏡頭並落在數字圖像傳感器上時，光線將被轉化成數字信號作為像素。數碼相機處理這些像素的時間差很小，但時間上的區別仍然存在。因此，這種區別創造了單個像素作為單個個體的概念。

當一顆正在融化的冰球被置於黑盒中，在穩定光源的照射下，這一過程會經歷多少種視覺變化呢？答案應該是無窮無盡且難以預知。若有一台能夠預見未來並展示所有可能性的機器，這個答案就會變得具體。同時，我們又該如何確保機器中數據的自主性呢？

When light travels through a lens and lands on the digital image sensor, that light will transform into digital signals as pixels. The digital camera processes those pixels in a minimal time difference, but the time distinction still exists; thus, the contrast creates a concept of a pixel as an individual.

How many visual variations will occur when a melting ice ball is placed inside a black box and exposed to a stable light source? The answer should be infinite and unpredictable. If there were a machine that could predict the future and show all the possibilities, then the answer would become concrete. In the meantime, how should the pixel's autonomy be maintained?

《Stochastic Camera》是一個自2018年起的藝術項目系列。《Stochastic Camera (version 0.2) - the melting crystal ball》對機器學習圖像處理的算法邏輯進行了檢視和反思。此項目從紀錄不同冰球在3小時內的融化開始，持續8天，拍攝了約24小時的冰球融化畫面，之後將其轉換為原始像素數據以進行進一步計算。融化的冰球象徵著未來或許不再需要用於預測的水晶球，同時它也作為一個折射物來延伸相機的鏡頭改變折射率。《Stochastic Camera (version 0.2) - the melting crystal ball》利用兩條算式，分別代表整體影像的穩定及個體像素自主。算式的靈感來自集合平均和隨機過程，計算數據的總平均值並預測每個像素的動量，最終形成兩組像素數據和移動圖像。

"Stochastic Camera" is a continuous art project that began in 2018. In version 0.2, re-examines and rethinks machine learning image processing algorithmic logic. The project starts with recording the melting of different ice balls over three hours. It continues for eight days, capturing approximately 24 hours of ice ball melting footage and converting it into raw pixel data for further processing. The melting ice ball symbolises a future world where crystal balls used for divination may no longer be needed while also existing as a refractive object to extend the camera's lens. The "Stochastic Camera (version 0.2)" utilises two independent formulas representing overall stability and individual autonomy. These formulas are inspired by ensemble averaging and stochastic processes, calculating the data's overall mean and predicting each pixel's momentum, which results in two sets of pixel data and moving images.

在典型的機器學習圖像處理流程中，特徵提取是讓機器從原始像素數據中獲得最相關信息的過程。這種方法將產生顏色、形狀、紋理等特徵集，有助於計算機科學家降低數據維度和考慮的隨機變量數量。神經網絡，如PixelRNN，也會將相鄰像素的顏色數據作為目標像素的子採樣圖像像素。這導致了一種有效且成本較低的方式來預測和生成圖像。然而，從單個像素的角度來看，作為生活在二維圖像中的個體，計算機科學家確實犧牲了單個像素的自主性，以減少計算時間和成本。在此一問，會不會有一種可以平衡自主性及效率的方法？

In a typical image processing workflow, feature extraction is the process that allows machines to identify the most relevant information from raw pixel data. It generates sets of features such as colour, shape and texture. These allow computer scientists to reduce the data dimensions and the number of random variables that need to be considered. Neural networks, such as PixelRNN, also consider the colour data of neighbouring pixels as subsampled image pixels for the target pixel. This results in an efficient, less time-consuming, and less costly way of predicting and generating image emergence. However, from the perspective of a single pixel, an individual living in a two-dimensional image, computer scientists sacrifice the autonomy of individual pixels to reduce computational time and cost.

演算法﹙一﹚：集合平均法此方法針對8個冰球的像素數據矩陣進行平均運算。透過該算法，各像素間存在著互相影響的關係，從而構成一個穩定的整體圖像。最終生產的圖像類似於將8個視頻片段疊加的效果，但在數據處理上具有顯著差異。該方法意味著通過將這些數據壓縮至單一平均值，消除所有個體數據的劇烈變化，將每一個像素綁定至一個整體代表。

演算法﹙二﹚：隨機過程 - 個體像素動量此方法先確定了單個像素高於或低於其平均值的概率。與其他用於圖像處理的神經網絡不同，此方法不借用鄰近像素數據作為參考來影響目標像素的變化。相反地，相鄰的像素數據完全不予考慮，賦予每個像素最大程度的自主性來進行演化湧現結果。

《Stochastic Camera (version 0.2) - the melting crystal ball》入選香港藝術中心第28屆ifva獎媒體藝術組決賽。

Algorithm 1: Ensemble Averaging

This method averages the pixel data arrays of the eight ice-ball shots. This formula shows a relationship between the pixels influencing each pixel to form a stable overall image. The final image is similar to superimposing 8 video clips but differs significantly in data handling. This approach means that all strong variations are eliminated by compressing this data into a mean value, and the overall average suppresses each pixel.

Algorithm 2: Individual Pixel Momentum

This method determines the probability of individual pixels being above or below their mean. Unlike other neural networks used for image processing, it does not use neighbouring pixel data as a reference to influence changes in the target pixel. Instead, neighbouring pixel data is not considered, and each pixel is given maximum autonomy to emerge.

"Stochastic Camera (version 0.2) - the melting crystal ball" has been a finalist in the Media Art category of the 28th ifva Awards at the Hong Kong Arts Centre.