Plant Leaf Recognition

Research on automatic leaf classification has been active since 2000. Lots of hand-crafted features have been proposed, ranging from shape based, to statistical texture and margin related [2] [3] [1]. Also generic computer vision object recognition features, such SIFT[32] and HOG[33], are studied for this problem. Most of such manually engineered features achieve excellent accuracy on clean images taken in controlled conditions, which consist of one single well aligned leave on contrasting background, such as those images in data set [15]. Recently, with the huge success of deep ConvNets, particularly from the winners of ILSVRC [11] [23], [24] and [22], researchers start to apply deep ConvNets to this problem. In [36], Sharif, et al. suggested that generic features can be extracted from large ConvNet and yield very good results on fine-grained classification problems even without fine-tuning the pre-trained model. [34] compares traditional approaches with ConvNet based approach, and discuss impacts on various conditions, including translation, rotation, scale and occlusion. However it only give results on Flavia[16] dataset which is relatively less challengingIII. [35] proposes a VGG[23] based architecture, and use multiple organ features. It produces above 70% mean average precision on ImageCLEF dataset[17]. III. DATASET