This paper proposes a multiscale conditional random field (mCRF) approach for image labeling, which incorporates contextual features at different scales to assign labels to each pixel in an image. The model combines the outputs of multiple components that differ in the information they encode and their scale. Some components focus on the image-label mapping, while others focus on patterns within the label field. Components also differ in their scale, with some focusing on fine-resolution patterns and others on coarser, more global structures. A supervised version of the contrastive divergence algorithm is used to learn these features from labeled image data. The mCRF model defines a conditional distribution over the label field given an input image by multiplicatively combining component conditional distributions that capture statistical structure at different spatial scales. The model includes three components: a local classifier, regional features, and global features. The local classifier uses a statistical classifier, such as a neural network, to classify pixels based on local image statistics. Regional features capture local geometric relationships between objects, while global features capture coarser patterns in the label field. The model is trained using a discriminative approach based on the Conditional Maximum Likelihood (CML) criterion, which maximizes the log conditional likelihood. The model is evaluated on two real-world image databases, the Corel image database and the Sowerby Image Database. The results show that the mCRF model outperforms a Markov random field (MRF) and a local classifier in terms of classification accuracy. The model's performance is attributed to its ability to capture both local and global relationships in the label field, as well as its direct representation of large-scale interactions. The paper also discusses the advantages of discriminative over generative modeling and the limitations of local interactions captured by MRF models. The mCRF model is shown to generate more reasonable labelings by incorporating contextual information from regional and global features, which corrects many of the wrong predictions from the local classifier. The model's performance is further improved by allowing label features to access image statistics. The paper concludes that the mCRF model provides a novel probabilistic framework for image labeling that effectively combines local and global information to produce accurate and consistent labelings.This paper proposes a multiscale conditional random field (mCRF) approach for image labeling, which incorporates contextual features at different scales to assign labels to each pixel in an image. The model combines the outputs of multiple components that differ in the information they encode and their scale. Some components focus on the image-label mapping, while others focus on patterns within the label field. Components also differ in their scale, with some focusing on fine-resolution patterns and others on coarser, more global structures. A supervised version of the contrastive divergence algorithm is used to learn these features from labeled image data. The mCRF model defines a conditional distribution over the label field given an input image by multiplicatively combining component conditional distributions that capture statistical structure at different spatial scales. The model includes three components: a local classifier, regional features, and global features. The local classifier uses a statistical classifier, such as a neural network, to classify pixels based on local image statistics. Regional features capture local geometric relationships between objects, while global features capture coarser patterns in the label field. The model is trained using a discriminative approach based on the Conditional Maximum Likelihood (CML) criterion, which maximizes the log conditional likelihood. The model is evaluated on two real-world image databases, the Corel image database and the Sowerby Image Database. The results show that the mCRF model outperforms a Markov random field (MRF) and a local classifier in terms of classification accuracy. The model's performance is attributed to its ability to capture both local and global relationships in the label field, as well as its direct representation of large-scale interactions. The paper also discusses the advantages of discriminative over generative modeling and the limitations of local interactions captured by MRF models. The mCRF model is shown to generate more reasonable labelings by incorporating contextual information from regional and global features, which corrects many of the wrong predictions from the local classifier. The model's performance is further improved by allowing label features to access image statistics. The paper concludes that the mCRF model provides a novel probabilistic framework for image labeling that effectively combines local and global information to produce accurate and consistent labelings.