Applying a multiple instance learning technique to image classication

n binary Multiple Instance Learning (MIL) the objective is to discrimi- Sept. 11th 16.30-18.30 Majorana nate between positive and negative sets of points. In the MIL terminology each set is called bag and the points inside each bag are called instances. In the case of two classes of instances (positive and negative), a bag is positive when it contains at least a positive instance and it is negative if it contains only negative instances. For such kind of problems there exist in literature two different approaches (see Amores 2013 and Carbonneau et al. 2018): the bag-level approach and the instance level approach. While in the former the total entity of each bag is considered, in the latter a classifier is obtained on the basis of the characteristics of the instances, without looking at the whole entity of each bag. We have applied to image classification an instance-level technique based on the Lagrangian relaxation of a Support Vector Machine (SVM) type model and we re- port some numerical results. In particular, given a set of images, for each of them we have performed a segmentation in the following way. Starting from a bitmap image, we have obtained an indexed image, where each element corresponds to a pixel and contains a triplet representing the RGB (red, green, blue) scale. Once the indexed image has been generated, the successive step has consisted in converting each indexed image (and the corresponding colormap) into a RGB image. After that, we have proceeded by grouping the pixels in square subregions of appropriate dimension: each image subregion forms the so called blob. In the MIL framework the images constitute the bags and the blobs correspond to the instances.