Research On Deep Learning Based Plant Species Identification

Plants play a really important role in human life because they are source of food,clothes and medicines.As huge amounts of plant species exists in natural world,how to recognize them preciously became a really challenge problem.In the past,people can only recognize unknown plant species by comparing them with pre-collected specimen or using a professional book.However,it is not convenient to carry multiple books especially when you need to take a long time travel.With the rapid development of key-word search engines,people can get sufficient information for object plant species by simply entering the name of object plant.However,for people without professional training,finding appreciate words to describe object plant is still a very difficult problem.Recently with the rapid development of image based classification technique,people can simply using a mobile phone to take a picture of object plant and get detailed introduction.There are many researches on image based plant species identification in recent few years.However,most of the existing researches focus on leaf-based plant species identification,and only deal with small scale plant species,which may not meet the requirement of real life.Very strong inter-species relations are existed in plant world,but such relations are completely ignored in softmax classifier,which means the network structure may not be able to achieve the global optimum during the training process.Besides,most of the deep networks structures focus on 1000 classes classification task,which may not perform well when recognizing large-scale plant species.Also when applying deep network structure on mobile devices,highly demand on computation speed and memory cost are required so most of the traditional deep network structures can not be directly used in mobile devices.1.Considering no existing suitable large-scale plant species dataset for training reliable deep network,we built a plant dataset called Orchid2608.The Orchid2608 containing 2608 plant species under 158 genera.Orchid2608 contain over 2000000 images for training and validation,and 180000 images for testing.Images for Orchid2608 dataset are originally clawed from multiple key-word search engines such as Bing,Google and Flickr.The duplicate images are removed automatically by clustering method.Finally,manually decoration method is used to improve the reliability of Orchid2608.Compared with existing plant dataset such as Flower101 or Plant Clef2016,Orchid2608 has larger scale and more suitable for a fine grained classification problem.2.Considering strong inter-species relations exists in plant world,a hierarchical multi-task learning method is proposed.A two-layer fixed plant taxonomy is constructed to organize large scale plant species hierarchically,so the visually similar plant species will be assigned into same learning task,and the gradients of loss function to each plant specie can be more uniform.More specific deep feature and classifier are learned for each learning task.Attention mechanism is also used to decrease the role of the highlevel common feature components in low-level classification task.During the training process,a specific path based training method is proposed to remove the error propagation effect in hierarchical structure.The experiments proved the proposed method can achieve competitive performance on large-scale plant species identification.3.Considering most of the existing deep network structure originally designed for recognizing 1000 object classes,a deep fusion method is also proposed to deal with largescale plant species by leverage the successful design of existing network.Plant taxonomy is used as a guideline for generating divers but overlapped task groups,and each task group should contain no more than 1000 plant species.Also a specific “and one” specie is added in each base deep network,and plant species not belonging to species in corresponding task group should be classified into the “and one” category.Finally,multiple outputs from all base deep networks are fused to form a high-level feature presentation and give the final prediction.The experiments shown that the proposed deep fusion network can achieve outstanding performance on large-scale plant species identification.4.Considering apply deep network structure on mobile devices,an improved light weight deep network is proposed.The channel shuffle operation which exists in Shufflenet and Shuffle Net V2 bring extra time costs while no extra FLOPs.So an improved bottleneck structure is proposed to achieve improvements on both computation speed and accuracy.Also the inter-species relations are also considered during the training process by proposing a specific simplified tree classifier and hierarchical orthogonal loss.The experiment results indicated that proposed light weight deep network structure can achieve competitive performance on both computation speed and final classification accuracy.