Multi-Class Recognition using Noisy Training Data with a Self-Learning Approach

Exploiting ConvNets for object classification systems requires extensive labor work, since these networks require to be trained by sufficiently large and accurately labeled datasets. We propose a novel self-learning approach, which is able to generate a reliable multi-class object classification model from a low-quality dataset that is disturbed with a high level of inter-class noise samples. This approach iteratively purifies the noisy training datasets for each class and updates the classification model. The iterations continue until the model and its parameters reach sufficient quality. The self-learning approach based on ConvNets is evaluated for a maritime surveillance use case, where vessels need to be classified into eight different types. The experimental results on the evaluation dataset show that the proposed approach improves the F1 score approximately by 5%, 8% and 25% at the end of the third iteration, while the initial training datasets contain 40%, 50% and 60% inter-class noise samples (erroneously classified labels of vessels), respectively. Additionally, the purification performance is highly dependent on inter- and inter-class similarities between training samples for higher noise levels. It was also found that the mean Average Precision (mAP) does not degrade so much, whereas other performance parameters show larger variation.