Representation Learning Framework of Object Recognition via Feature Construction

In this paper, we recognize objects within images by collecting information from a large number of random-size patches of the image. The different backgrounds accompany the foreground object demand to have a learning system to identify each patch as belonging to the object category or to the background category. We strengthen a recent method called Evolution-COnstructed (ECO), which is based on the ensemble learning approach which combines several weak classifier. The improvement is relying on decreasing the overfitting problem. Two different improving ideas are proposed: 1) Pooling operation, which is applied to the weak classifiers data, 2) Random Forest algorithm, which combines the weak classifiers outcomes. Experimental results are reported for classification of 9 categories of Caltech-101 data sets and proved that our modifications boost the performance over the base method and other existing methods.