| Plants are the most diverse organisms on Earth.They are widely distributed on land,rivers,lakes and oceans,and play an irreplaceable role in the ecosystems,material cycles and energy flows of the biosphere,where they are most critical.Understanding the species,properties and functions of plants helps people to modify the environment,make medicines,develop agriculture and so on.Therefore,knowing plants is not only necessary for botanists to conduct research on plant species diversity,but also for ordinary people to get close to nature and protect ecology,and it is an important part of the science curriculum for comprehensive scientific literacy enhancement.However,the identification of a large number of plants that are both different and similar to each other requires specialized knowledge of the plant classification system and various characteristics of plants such as stems,roots,fruits,leaves,and flowers,which undoubtedly greatly increases the difficulty of plant recognition for most ordinary people.Compared with other parts of plants,the shape and structure of plant leaves are stable and their images are the easiest to capture.Therefore,using image processing technology to identify the correct species of plants through leaf feature extraction and comparison,and then tell users the classification results,is undoubtedly the best path to carry out plant knowledge popularization.Traditional plant leaf feature extraction mainly focuses on color,shape,size,texture,etc.,but in fact,plant leaves have complex variation,and there are many small differences between leaves of the same category,and leaves of different species may also have very similar shapes or textures.Therefore,the classification based on plant leaf images is very challenging.With the emergence of deep learning,convolutional neural networks have shown excellent performance in image classification tasks.ResNet is currently the most popular image classification neural network and is widely used in various image classification subfields,but the underlying network model of ResNet does not capture the fine-grained features among plant leaf classes well,and the proposed ResNet-based improved Res2Net model,which is based on the improved ResNet,can better extract the features at different scales.Meanwhile,most of the current plant leaf classifications increase the diversity of features to improve the performance of the model by data augmentation,while multi-scale feature extraction can also increase the diversity of features.In summary,this paper will investigate the application of convolutional neural network on plant leaf image classification.In this paper,two methods of plant leaf image classification based on convolutional neural networks are proposed for small plant leaf datasets,and the main research contents are as follows.1.A classification algorithm for real small plant leaf image datasets is proposed for the problem of recognition difficulties of traditional convolutional neural networks for fine-grained leaf classification among plant leaf classes.The algorithm proposes a new algorithmic model DMSNet(Deep Multi-Scale Network)based on the multi-scale skeleton structure model Res2Net-50 with the addition of multi-scale inputs,which is more suitable for plant leaf image classification by increasing the diversity of features to allow the model to acquire fine-grained features from different perspectives.2.In response to the problem that the DMSNet model does not pay enough attention to the fine-grained features of plant leaves,a new attention mechanism ECA module is introduced into the original module Basic2 Block of Res2Net network for improvement,and ADMSNet(Attention-based Deep Multi-Scale Network)is proposed.The improved module is called ECA-B module,and all the Basic2 Block used in DMSNet is replaced by the new ECA-B module in this paper.The comparison experimental section compares the initialization models of DMSNet and ADMSNet with the traditional and commonly used initialization models of convolutional neural networks such as ResNet and Res2Net.The experimental results show that both the DMSNet and ADMSNet algorithm models proposed in this paper outperform the traditional convolutional neural network algorithms.In the ablation experiments,all the Basic2 Block modules of Res2Net-50 are replaced with ECA-B modules for comparison with ADMSNet,which further proves that the multi-scale input can well improve the performance of the convolutional neural network model for plant leaf image classification.All experimental results show that the application of DMSNet and ADMSNet algorithm models on small plant leaf image datasets is feasible,and they can effectively solve the fine-grained leaf classification problem.3.The ECSANet(Efficient Channel and Spatial Attention Network)algorithm is proposed for the small plant leaf image classification problem.The main innovative work of this algorithm makes full use of attention information while achieving a lightweight model by combining spatial and channel attention.The classification performance of the convolutional neural network model for plant leaves is improved.The experimental part compares ECSA with the commonly used attention module on small plant leaf image dataset,and the experimental results show that the proposed ECSA module in this paper outperforms the traditional attention mechanism module on small plant leaf dataset.To further demonstrate the generalization of the ECSA module,experimental comparisons are also performed on the public dataset Cifar-10,and the experimental results show that the ECSA module also has some performance improvement on the backbone network model on other datasets.All the algorithms and models proposed in this paper have achieved better results in several groups of comparison experiments,which prove that the methods proposed in this paper are feasible to be applied in plant leaf image datasets,and all of them can effectively solve the problem of identification difficulties in fine-grained leaf classification among plant leaf classes. |
PDF Full Text Request