Improving Model Performance for Plant Image Classification With Filtered Noisy Images

The training of convolutional neural networks for image recognition usually requires large image datasets to produce favorable results. Those large datasets can be acquired by web crawlers that accumulate images based on keywords. Due to the nature of data in the web, these image sets display a broad variation of qualities across the contained items. In this work, a filtering approach for noisy datasets is proposed, utilizing a smaller trusted dataset. Hereby a convolutional neural network is trained on the trusted dataset and then used to construct a filtered subset from the noisy datasets. The methods described in this paper were applied to plant image classification and the created models have been submitted to the PlantCLEF 2017 competition.