*5.5. Comparison of Di*ff*erent Methods*

We also compared the proposed method with other methods on some datasets. To test the effectiveness and extensibility of the proposed feature and system, leaf datasets Flavia, ICL [37], Swedish [38,42], and MEW2012 [32] were used for testing.

As ICL contains so many leaf images, most methods always take a part of ICL dataset for testing. To compare with the MEW method [32], we followed its setting. On each dataset, for each species, half of leaf images were chosen as training sets, and the rest were the testing set. Supposing the number of species is *p*; if *p* is an even number, the training leaf images number was *p*/2; otherwise, the training leaf images number was (*p* + 1)/2. Finally, the training set and the testing set were roughly equal (in fact, the testing set was larger than the training set). The detailed data of the four datasets are shown in Table 2. In the following, all tests were repeated 10 times to get a convincing result.

**Table 2.** Detail information of the four datasets.


First, we compared these methods on the Flavia dataset. The comparison results with other methods are shown in Table 3. ZRM [41] is a method based on Zernike moments. Z&H represents the method of Ref. [11], which is based on Zernike moments and histogram of oriented. VGG16 [42] and VGG19 [42] are the pre-trained models based on CNN architecture with logistic regression. MLAB (Margin, lobes, apex and base) [43] is the phenetic features of leaf. MLBP [44] is the method of extracting texture features based on modified local binary patterns. Muammer Turkoglu and Davut Hanbay [45] proposed the improved descriptors based on LBP, called region mean-LBP (RM-LBP), overall mean-LBP (OM-LBP), and ROM-LBP. RIWD (rotation invariant wavelet descriptor) [46] is a new shape proposed by Ehsan Yousefi et al. GIST [47] is an approach for plant recognition using GIST texture features. Wang et al. [48] proposed a few-shot learning method based on the Siamese network framework (S-Inception) to better classify the small sample size (where *n* is the number of species used in this experiment and the number of trainings is 20 *n*). Most of these comparison methods do not introduce the number of training and test samples. Among the comparative methods, the deep learning-based method [42,48] does not obtain the best recognition results, but is slightly lower than other machine learning methods [44–46]. SSV [17] is a fusion feature composed of 11 shape features, 7 statistical features, and 5 vein features. The recognition result of SSV is slightly higher than our proposed method. As shown in Table 3, the training samples of the experiment are far more than the test sample images. It can be seen from the method Z&H [11] in Table 4 that, when the number of training samples increases and the number of test samples decreases, the recognition rate increases. Further, our method uses more total images than SSV. The total images of SSV were 1600, while our total images were 1907. More than 300 images were removed in SSV. The Flavia dataset we used is original and unfiltered. Therefore, it is understandable that the SSV method obtains a slightly better

recognition rate under the very superior experimental conditions. Overall, the proposed method is superior to most of other existing methods.


**Table 3.** Comparison of proposed method with existing methods on Flavia dataset.

**Table 4.** Comparison of proposed method with existing methods on Swedish dataset.


Table 4 shows the results of different methods on the Swedish dataset. It contains 1125 sample images from 15 species, with 75 images per species. The authors of [49] proposed SMF, which utilizes the area ratio to quantify the convexity/concavity of each contour point at different scales to construct margin feature, and they used a combination of morphological features as shape feature. Yang et al. [50] introduced a novel multiscale Fourier descriptor (MF) based on triangular features, which effectively captures the local and global features of leaf shape. MARCH [51] (multiscale arch height) is a novel multiscale shape description. Wang et al. [52] proposed a hierarchical string cuts (HSCs) method. CSD [53] is a counting-based shape descriptor for leaf recognition, which can capture global and local shape information independently. CBOW is a shape recognition algorithm based on the curvature bag of words (CBOW) model. Generally, the recognition accuracy is improved with the increase of the number of training samples. When the training number of the method Z&H [11] is 750, the recognition result is significantly improved, which is slightly higher than the method we propose. In addition, compared with the other existing methods, the proposed method is superior. S-Inception [48] obtained the lowest recognition accuracy, while MEW [32], MF [50] and CSD [53] were close to the accuracy of the proposed method. The recognition accuracies of the other methods were also very close.

For ICL dataset, some researchers only use part of samples from dataset. Hence, the detailed comparisons are listed in Table 5. GTCLC [55] is a leaf classification method using multiple descriptors. Cem Kalyoncu et al. proposed a new local binary pattern (LBP) descriptor, and they combined it with geometric, shape, texture, and color features for leaf recognition. The authors of [56] used several different descriptors to extract texture and shape features and proposed a pre-training method based on the PID to improve the DBNs. DWSRC (discriminant WSRC) [57] is the method proposed by Zhang et al. for large-scale plant species recognition. The authors of [58] presented the novel relative sub-image sparse coefficient (RSSC) algorithm for mobile devices. DBNs chose 50 species for training and testing and it obtained the highest accuracy with 96%, higher than the proposed method; however,

when the number of species in the experiment was 220, the recognition accuracy dropped to 93.90%. When 220 species were selected for training and testing, our proposed method achieved the highest accuracy 94.22%.


**Table 5.** Comparison of proposed method with existing approaches on ICL dataset.

MEW dataset is also a large dataset. We compared our method with some classic methods, as shown in Table 6. The PCNN proposed by Wang et al. [59], based on pulse-coupled neural network and SVM, is a novel plant recognition method. PCNN and DPCNN have better performance than the others. It is obvious that the method we propose is better than the other methods.

**Table 6.** Comparison of the proposed method with existing approaches on MEW dataset.

