Quais são as técnicas de clustering de dados de imagem mais comuns?

Although K-means is very popular clustering algorithm and widely used, one needs to be cautious when using it. K-means can fail when clusters of arbitrary shapes are present in the data because of sub-optimal selection of representative data points and neglect of some clusters.

K-means clustering is a widely used image data clustering method. It begins by selecting a set number of cluster centers (k) and assigns each image to the nearest center using a distance measure like Euclidean distance. The algorithm then updates cluster centers by averaging images within clusters and repeats this process until convergence. K-means is known for its simplicity and speed, but it can be sensitive to initial center selection, the number of clusters, and outliers, which are limitations to consider.

K-Means has been shown to be less accurate than traditional object detection methods such as Scale-Invariant Feature Transform (SIFT), Speeded-Up Robust Feature (SURF), Histogram of Oriented Gradient (HOG), and Deep Learning-based object detection. However, it has a significant advantage in terms of computational efficiency. This trade-off between accuracy and efficiency needs to be considered when choosing a method for a specific application.

Uma das áreas mais populares para aprendizado de máquina é a Visão Computacional, onde processamento de imagens e IA se combinam. Métodos de clusterização detectam padrões e agrupamentos sem intervenção humana. K-Means, DBSCAN, Spectral Clustering e Clustering Hierárquico são as principais técnicas aplicadas, dependendo da natureza dos dados e dos requisitos específicos do problema de clusterização em imagens.

A Good painting or photograph is defined by center point, balance and composition. The composition defines the curve how center points are connected e.g. S shape. Similarly, K-means clustering search for center points and balance when used within a photo/painting. Center points are defined by centroid and they should be far away from each other. The distance is calculated by Euclidean distance. This is how we can segment an image. K-means clustering algorithm can also find similar images when used across images. K-means clustering algorithm uses visual similarity for finding similar images. K-means algorithm is sensitive to noise, so we should remove noise and improve image quality before clustering.

In image data clustering, hierarchical techniques are like zooming in and out on a picture. They let you explore the granularity of image similarities, from broad strokes to fine details. Agglomerative clustering is akin to building a mosaic—starting with individual tiles (images) and combining them into a cohesive picture (clusters). Divisive is the reverse, like starting with a full image and finding the lines that divide it into distinct segments. This method paints a clear hierarchy of image relations but can be costly, requiring a careful choice of how you measure closeness—whether you're viewing the picture as a whole or focusing on individual pixels.

Assume you are living in a exotic location near the beach. You admire nature, so you take photograph and want to cluster them. The photo would be clustered on the basis of texture of sky and reflection of sky in ocean. Initially, you will get only two clusters, one for day and other for night. It will be further clustered on the basis of time of the day. Like K-means clustering, hierarchical clustering also uses distance measure e.g. Euclidean distance. It is computationally expensive, as you will get cluster in hierarchical nature. Hierarchical clustering can compare similarity between images in reverse order also, where individual images are compared together. In this way, you can start comparison from leaf nodes.

Hierarchical clustering is an unsupervised machine learning algorithm that groups similar data points together by building a hierarchy of clusters. It is a bottom-up approach, meaning that it starts with each data point as its own cluster and then merges clusters together until it reaches a desired number of clusters or a stopping criterion is met. Agglomerative clustering: This common hierarchical clustering approach begins with individual data points as clusters and iteratively merges the closest clusters until a set number or stopping criterion is achieved. Divisive clustering: data points initially form a single cluster, progressively dividing it into smaller clusters until a target number or stopping criterion is reached.

In English "Spectral" means "like a ghost" and ghost can change form very easily. Similarly, Spectral Clustering changes many forms during clustering process. It starts with building similarity matrix from images. Then Eigen values is computed from similarity matrix which carries latent features. Latent matrix is changed to adjacency matrix, so that it can present graph in the form of nodes and edges. Here, Edges carries similarity scores and which is later processed by traditional ML Algorithms like PCA, and k-means.

I find spectral clustering quite useful and powerful, in terms of outcome/resources trade-off. But you must keep an eye in complexity and scalability. Eigenvalue decomposition and the laplacian matrix can be computationally expensive. On the other hand, it may not suitable for very large datasets. Despite this weaknesses, its ability to discover clusters with complex and non linear shapes is actually straightforward, and it can be applied in various domains, being image segmentation a good candidate. Finally, the use of an eigenvalue problem connects it directly with maths, which I find as a mark of beauty.

In this Spectral clustering technique , that distinguishes clusters within datasets by analyzing the eigenvalues of a similarity matrix. This method is particularly valuable when working with data in high-dimensional spaces and is adept at uncovering intricate, non-linear structures in the data. Spectral clustering finds practical applications in diverse fields, such as image segmentation, social network analysis, and the detection of communities within networks.

Density-based clustering in image data is like recognizing crowds in a satellite photo. DBSCAN, one of the most renowned algorithms, picks out clusters by seeking areas where data points (people) gather closely—a high-density crowd—while distinguishing these from sparse areas (empty spaces). It's adept at ignoring noise (birds flying over) and doesn't need pre-set cluster numbers, much like not needing to know how many gatherings there are before looking at the photo. However, setting the right parameters is crucial—it's like adjusting the lens to ensure you don't mistake a tight-knit group for a random scattering or miss a small gathering altogether.

One of the most well-known density-based clustering algorithms is DBSCAN (Density-Based Spatial Clustering of Applications with Noise). This technique is particularly useful for applications where clusters are irregularly shaped and may have varying densities, making it suitable for tasks such as anomaly detection and spatial data analysis. An example of density-based clustering using the Let's say you have a dataset of customer locations in a city, and you want to group them into clusters for targeted marketing. DBSCAN would identify clusters based on the density of customers in specific areas. Customers located in a busy shopping district may form one cluster, while those in quieter neighborhoods form separate clusters.

In my experience, autoencoders work very well in this type of task. Designing the right network to compress each image into a salient, lower-dimensional vector is important. With information-rich vectors, the clustering step becomes less of a headache, and the approach has other applications far beyond computer vision too.

In a corporate set-up, if you have a access to cloud infrastructure it is always better to use object detection / computer vision API service offered by the cloud service provider and vectorise (kind-off) the features. Now you can apply any clustering method to test. This is easy, fast and especially helpful during proof of concept stage.