Automatic Image Annotation with Real World Noisy Data

Automatic Image annotation is an important open problem in computer vision. In real world dataset environment, image labels are often noisy. For the task of image annotation with weakly labels, we propose SNLWL, a semantic neighborhood learning model from weakly labeled dataset. Missing labels are replenished using reweighting the error loss function. Then semantic balanced neighborhood is construct for samples in the training set. The methods allows the integration of multiple label metric learning and local nonnegative sparse coding. In this manner, we can optimally construct semantic consistent neighborhood where neighbors have higher global similarity, partial correlation, conceptual similarity along with semantic balance for samples in the training set. We also introduce an iterative denoising method of the label predictions to handle the noise. We investigate the performance of different variants of our method and compare to existing work. We present experimental results for various data sets. On all datasets, SNLWL makes a marked improvement as compared to the current state-of-the-art.