Comparison of ResNet50 and MobileNetv2 Methods for Agricultural Crop Classification.pdf

Comparison of ResNet50 and MobileNetv2 Methods for Agricultural Crop Classification Abstract—Agriculture is an important sector in and analyze hyperparameters for plant type economic development. The application of technology, classification using two well-kn...

Comparison of ResNet50 and MobileNetv2 Methods for Agricultural Crop Classification Abstract—Agriculture is an important sector in and analyze hyperparameters for plant type economic development. The application of technology, classification using two well-known CNN models, especially machine learning, is needed to manage namely ResNet50 and MobileNetV2. agricultural data. This research aims to develop a The main objective of this research is to answer classification model by utilizing deep learning models, namely ResNet50 and MobileNetV2, for Four the question of how optimal hyperparameter settings agricultural crops: rice, corn, cassava, and sugarcane. in these two models can improve the accuracy of Both models were trained on the same plant dataset, plant type classification. This question is important and hyperparameters such as image size and number because accurate classification has broad of epochs were tuned. The research results show that applications in precision agriculture, crop proper hyperparameters can significantly improve monitoring, and technology-based agricultural accuracy. The ResNet50 and MobileNetV2 models management systems. achieve accuracies of 63.83% and 87.23%, This research is expected to make a significant respectively. This research proves that the use of deep contribution to the field of digital image recognition learning models and hyperparameter tuning can and plant type classification by providing an in-depth effectively improve plant type classification analysis of the influence of hyperparameters on the performance. performance of ResNet50 models and MobileNetV2. Keywords— classification, hyperparameters, The objective of this research is to deliver a more MobileNetv2, ResNet50 effective and efficient classification system, which will ultimately support innovation in the field of I. INTRODUCTION agricultural technology. Convolutional Neural Network (CNN) is a deep II. RELATED WORKS learning algorithm used to recognize digital images. As a digital image recognition algorithm, this Classification using two models derived from algorithm is very effective for classification and has CNN, such as ResNet50 and MobileNetV2, is often many models. Digital image recognition has used and compared regarding accuracy and great potential for facial recognition , pattern processing speed. However, no one has yet identification, and classification. The CNN structure classified several types of agricultural crops using consists of many layers, including convolution, this model. Classification using the ResNet50 and pooling, and fully connected layers. CNN works by MobileNetv2 algorithm models has been used, but extracting important features in input data, reducing the research objects or data used in this research are dimensions, and classifying data. four breeds of cattle endemic to Indonesia. CNN models that can accommodate integration Previous study examined medicinal plant needs into mobile applications include classification using the ResNet50 model, which MobileNetV2 and ResNet50. In CNN, ResNet50 utilizes the depth of the ResNet50 layers to be is a model that has higher accuracy in image trained on a comprehensive dataset of medicinal leaf recognition than other models, such as VGG16 and images and uses transfer learning technology. VGG19. The Renet50 model can effectively This study used 80 types of medicinal plants and the solve the problem of gradient disappearance and ResNet50 classification model was able to achieve network degradation by using connections in a deep accuracy of up to 99.84%. However, the data used to CNN model using a residual structure. train the model was only an image of a single leaf, MobileNet is a CNN model that is effectively so the model may not be able to recognize medicinal used as a classification model whose use is aimed at plants that are still alive or intact. mobile devices, especially the MobileNetV2 model. ResNet50 and GoogleNet were modified as Image testing results using MobileNetV2 have a Classification Similarity Network (CSN) for image higher level of accuracy and shorter training time fusion with medical images. In a study on than the MobileNetV1 model. The MobileNetV2 developing the ResNet50 model with the transfer model is ideal for implementation on mobile devices learning method used to speed up the development with limited resources. process, the results reached 73.33% for accuracy In the context of plant type classification, where the model implemented Python and selecting the right CNN model and optimizing TensorFlow. ResNet50 with SVM achieves an hyperparameters is very important to improve accuracy of up to 99% and ResNet50 using the accuracy and efficiency. Hyperparameters such Softmax model achieves an accuracy of 98.8%. as learning rate, image size, and number of epochs Another study shows that adding layers to the significantly influence model performance. Setting ResNet50 architecture can increase the accuracy of the right hyperparameters can increase model wheat leaf disease classification, reaching up to accuracy. This research aims to develop a model 98.44%. XXX-X-XXXX-XXXX-X/XX/$XX.00 ©20XX IEEE  ResNet50 was used to detect and classify rotten difficult due to the addition of attention modules fruit, resulting in a validation accuracy of 98.89%. with a classification time of around 0.2 seconds per In another study, the MobileNetV2-UNET image. Although the results are good on monotonous model was used to segment leaves of guava, potato backgrounds, accuracy decreases to 85% on images and jamun plants with an accuracy of 96.38%. with complex backgrounds. Although efficient However, its accuracy is still lower than that of the and fast ResNet50, thanks to its pre-trained hybrid Plant Species Detection Stacking Ensemble architecture, is not always suitable for all datasets, Deep Learning Model (PSD-SE-DLM), which utilizing ResNet50 with transfer learning can achieved 99.84% accuracy for plant classification provide an accuracy of 93.5%. The ResNet50. The MobileNetV2 model is used for the algorithm using the Cyclical Learning Rate produces classification of traditional Indian medicinal flowers a model with an accuracy of up to 95%. In this with the highest accuracy reaching 98.23%, showing research, the setup with a small CLR and no dropout its superiority in flower images that have different has produced good accuracy and loss results. environmental backgrounds. In another study, Classification using MobileNetV2 is very the MobileNetV2 model was tested with several effective, in a lotus plant classification study the pre- different parameters to achieve the best accuracy, trained MobileNetV2 model was able to classify producing an accuracy of 75% on training data and names in the lotus class with the highest accuracy of 67% on validation data in ornamental plant 99.5%. In another study, the MobileNetV2 classification. This result is better than the built model achieved stable accuracy of 95.75%, 96.74%, CNN model, even though the accuracy level is still and 96.23% on datasets 1, 2, and 3 respectively in relatively lower than expected. fruit image classification. MobileNetV2 with hyperparameters was used In addition, MobileNetV2, with its lightweight for grape leaf disease classification and achieved network architecture and inverted residual module 99.94% accuracy after hyperparameter optimization with linear bottleneck, is increasingly popular in using a hyperband strategy, showing superiority various deep learning applications due to its memory over traditional CNN models in computational efficiency and high accuracy. In the melanoma efficiency and memory usage. In another study, classification study, MobileNetV2 was used as a MobileNetV2 was used for seed classification and base model by adding a global pooling layer and two achieved 98% accuracy on training data and 95% on fully-connected layers, achieving an accuracy of up testing data, demonstrating the superiority of a to 85% on the ISIC-Archive dataset, showing its simple and efficient architecture in memory usage superiority over other models such as ResNet50V2 compared to traditional CNN models. and InceptionV3. In the fruit image classification study, MobileNetV2 was modified to TL-MobileNetV2 by adding five special layers at III. PROPOSED MODEL the head of the model. Using transfer learning, TL- A. Datasets MobileNetV2 achieves 99% accuracy, 3% higher The dataset was obtained from Kaggle with than the original MobileNetV2, and shows improved search categories in the form of images of four types performance compared to AlexNet, VGG16, of plants that are the subject of classification. The InceptionV3, and ResNet. MobileNetV2, images used are in jpg format and there are a total of designed for mobile applications, excels in accuracy 245 images, 80% of which are used for training and on fruit datasets without pre-processing techniques, 20% for testing. demonstrating the superiority of its architecture in computer vision tasks. In other research, MobileNetV2 was combined with the Convolutional Block Attention Module (CBAM) to form the CBAM-MobileNetV2 model, which was used to classify citrus huanglongbing disease. Using transfer learning and parameter fine- tuning, CBAM-MobileNetV2 achieves 98.75% accuracy, higher than the MobileNetV2, Xception, and InceptionV3 models. This study shows that combining CBAM with MobileNetV2 improves the Fig 1. Images of Rice, Corn, Cassava, and Sugarcane model’s ability to capture semantic information and improves classification accuracy. CBAM- B. ResNet50 Model Architecture MobileNetV2 has the disadvantages of higher model complexity and longer training time compared to the ResNet50 was introduced by Microsoft Research original MobileNetV2, which also leads to greater in 2015 and won the ImageNet image recognition memory usage. In addition, the risk of overfitting competition which demonstrated the superiority of increases and implementation becomes more ResNet in image recognition. This ResNet model consists of several Residual blocks that overcome the problem of degradation in deep networks by using shortcut connections. The model receives an input image of some size in pixels, which goes through several stages of convolution, batch Fig 3. MobileNetV2 Architecture normalization, and ReLU activation. Residual blocks include identity blocks and convolutional blocks with multilevel filters (64, 128, 256). D. Hyperparameters Shortcut connections add together the original input Hyperparameters are used for the learning or with the convolution output, helping to retain testing process with two types of parameters information. After going through pooling and including image size and epoch. There are several flattening layers, the model uses Dense and Dropout image sizes tested including 32x32, 50x50, 75x75, layers before producing output with a softmax 100x100, 224x224, and 448x448 pixels. The activation function for multi-class classification. training process is carried out during different The model was compiled with the Adam epochs, including 50 epochs, 100 epochs, and 150 optimizer and categorical crossentropy loss, then epochs. trained for 150 epochs. Testing data is used for Each combination of image size and number of evaluation, measuring accuracy, precision, recall, f1 epochs is tested to see its effect on model score, and creating a confusion matrix. Training and performance. The evaluation results of each evaluation results are saved in text files, while hyperparameter combination are compared using accuracy and loss graphs are visualized for metrics such as accuracy, precision, recall, and f1 monitoring model performance. score. The best performance is recorded and used to determine the optimal hyperparameter settings that provide the most accurate and efficient results. IV. RESULTS AND DISCUSSION Fig 2. ResNet50 Architecture Based on the test results, it can be seen that image size significantly impacts the performance of C. MobileNetV2 Model Architecture both models. Larger image sizes tend to provide MobileNetV2 was proposed as a model for the higher accuracy. It can be seen on MobileNetV2 first time by the Google Research team in 2019 as an with an image size of 448x448, achieving the improvisation of the MobileNetV2 model to highest accuracy of 87.23% with 50 and 100 epochs. effectively maximize accuracy by considering Fig 4 shows the testing result of MobileNetV2. resource limitations on mobile devices. This MobileNetV2 model uses additional layers for classification, such as GlobalAveragePooling2D, Dropout, and Dense, which ends with a softmax layer for class prediction. The base layers of MobileNetV2 are frozen to prevent updates during training, while additional layers are trained using the training data. The model is compiled with the Adam optimizer and sparse categorical crossentropy loss. The training process was carried out for 150 epochs with a batch size of 32. Training and testing data were normalized to the range [0, 1] before use. After training, the model is evaluated using test data to calculate metrics such as accuracy, precision, recall, Fig 4. Testing result of MobileNetV2 and f1 score. Training and evaluation results are saved in text files, while accuracy and loss graphs On ResNet50, an image size of 75x75 provides are visualized to monitor model performance. The better results than other image sizes, with accuracy trained model is also saved in ‘keras’ format. reaching 63.83% at 150 epochs. However, the 100x100 image size shows a decrease in performance at higher epochs, indicating that ResNet50 may not manage higher image resolutions effectively in this setting. Fig 5 shows testing result of ResNet50.  that this model has a better balance between precision and recall. C. Comparison between MobileNetV2 and ResNet50 Overall, MobileNetV2 shows better performance than ResNet50 in image classification tasks on this dataset. Apart from accuracy, MobileNetV2 also excels in terms of precision, Fig 5. Testing result of ResNet50 recall and F1 score. MobileNetV2 achieved the highest accuracy of 87.23%, the highest precision of 89.29%, the highest A. Effect of the Number of Epochs on Model recall of 87.09%, and the highest F1 score of 87.61% Performance on images measuring 448x448 with 50 epochs. The number of epochs also has a significant Meanwhile, ResNet50 achieved the highest influence on model performance. In general, accuracy of 63.83%, the highest precision of increasing the number of epochs tends to increase 71.55%, the highest recall of 64.29%, and the model accuracy up to a certain point, after which highest F1 score of 59.44% on images measuring accuracy can decrease or stabilize. For example, on 75x75 with 150 epochs. MobileNetV2 with an image size of 224x224, the MobileNetV2 also shows better performance highest accuracy was achieved at 50 epochs stability at larger image sizes and higher number of (78.72%) and tended to be stable at 100 and 150 epochs, compared to ResNet50 which tends to show epochs. performance fluctuations. On ResNet50, increasing the number of epochs from 50 to 100 and 150 shows a general increase in D. Conclusion accuracy, but the best results remain at an image size of 75x75 with 150 epochs. Based on the test and analysis results, it can be concluded that MobileNetV2 is superior to ResNet50 in image classification tasks on this dataset. Apart B. Effect of Image Size and Number of Epochs on from accuracy, MobileNetV2 also shows better Precision, Recall, and F1 Score performance in terms of precision, recall and F1 score. Larger image sizes and optimal number of Apart from accuracy, precision, recall, and F1 epochs make a positive contribution to improving score also provide important insights into model model performance. This research provides valuable performance. Precision indicates the proportion of insights into the selection of model architecture and positive predictions that are truly positive. hyperparameter settings for image classification MobileNetV2 achieved the highest precision of tasks. 89.29% on an image size of 448x448 with 50 epochs, while ResNet50 achieved the highest V. EVALUATION precision of 71.55% on an image size of 75x75 with Evaluation using the Confusion matrix provides 50 epochs. The higher precision in MobileNetV2 a detailed picture of the model’s classification indicates that this model is more effective in performance by showing the number of correct and reducing false positive predictions. incorrect predictions made for each class. The Recall measures the proportion of total positive following is a confusion matrix analysis for data that is actually detected by the model. MobileNetV2 and ResNet50 based on their MobileNetV2 achieved the highest recall of 87.09% respective best configurations. on an image size of 448x448 with 50 epochs, while ResNet50 achieved the highest recall of 64.29% on an image size of 75x75 with 150 epochs. The higher recall on MobileNetV2 indicates that this model is better at detecting all positive data in the dataset. F1 Score is the harmonic mean of precision and recall, which provides a comprehensive picture of a b the balance between precision and recall. MobileNetV2 achieved the highest F1 score of Fig 6. Confusion matrix 87.61% on an image size of 448x448 with 50 a. 448x448 image size and 50 epoch epochs, while ResNet50 achieved the highest F1 b. 75x75 image size and 150 epoch score of 59.44% on an image size of 75x75 with 150 epochs. The higher F1 score on MobileNetV2 shows This confusion matrix analysis was carried out for MobileNetV2 and ResNet50 based on their respective best configurations. In the confusion matrix there are four key values: True Positive (TP), larger and more diverse datasets is also False Positive (FP), False Negative (FN), and True recommended, as well as investigating other models Negative (TN). The purpose of the evaluation in this such as EfficientNet and DenseNet to evaluate study was to measure the performance of two potential performance improvements in this different architectural models, ResNet50 and classification task. By following these MobileNetV2, in plant image classification by recommendations, it is hoped that improvements in varying the image size parameters (32x32, 50x50, model performance in image classification can be 75x75, 100x100, 224x224, and 448x448 pixels) and achieved, as well as better contributions to research the number of epochs (50, 100, and 150). This test and practical applications in the fields of image monitors accuracy, precision, recall and F1 score recognition and computer vision. using the formula: 𝑇𝑃 + 𝑇𝑁 ACKNOWLEDGMENT 𝐴𝑐𝑐𝑢𝑟𝑎𝑐𝑦 = 𝑇𝑃 + 𝑇𝑁 + 𝐹𝑃 + 𝐹𝑁 Thanks to research and community service 𝑇𝑃 institution, Universitas Pendidikan Ganesha for the 𝑃𝑟𝑒𝑐𝑖𝑠𝑖𝑜𝑛 = fund given. And also, KGEO Thailand for the 𝑇𝑃 + 𝐹𝑃 𝑇𝑃 research collaboration. 𝑅𝑒𝑐𝑎𝑙𝑙 = 𝑇𝑃 + 𝐹𝑁 2 × 𝑅𝑒𝑐𝑎𝑙𝑙 × 𝑃𝑟𝑒𝑐𝑖𝑠𝑖𝑜𝑛 REFERENCES 𝐹1 𝑠𝑐𝑜𝑟𝑒 = Sushma, L. and Laksmi, K. P. “An Analysis of Convolution 𝑅𝑒𝑐𝑎𝑙𝑙 + 𝑃𝑟𝑒𝑐𝑖𝑠𝑖𝑜𝑛 Neural Network for Image Classification using Different Models”, International Journal of Engineering Research Performance comparison between & Technology, vol. 9, no. 10, pp: 629-637, 2020. doi: MobileNetV2 and ResNet50 shows that https://www.ijert.org/research/an-analysis-of- MobileNetV2 has better accuracy and balance convolution-neural-network-for-image-classification- using-different-models-IJERTV9IS100294.pdf between precision and recall. The larger image size Budiman, A., Fabian, Yaputera, R. A., Achmad, S., and and optimal number of epochs in MobileNetV2 Kurniawan, A., “Student attendance with face recognition provide better results compared to ResNet50. The (LBPH or CNN): Systematic Literatur Rewiew”, Elsevier, results of this study confirm that MobileNetV2 is a vol. 216, pp. 31-38, 2023, doi: 10.1016/j.procs.2022.12.108. better choice for image classification tasks on this Fikri, A., and Murni, A., “New Generation Indonesian dataset. For practical implementation, it is Endemic Cattle Classification MobileNetV2 and recommended to use MobileNetV2 with a larger ResNet50”, Jurnal Ilmiah Teknik Elektro Komputer dan image size and a sufficient number of epochs to Informatika, vol. 9, no. 4, pp. 941-950, 2023. doi:10.26555/jiteki.v9i4.26659 achieve optimal accuracy and balance. The Astuti, R. P., Rachmawati, E., Siregar, S., Sardi, I. L., Fahmi, conclusion of the evaluation using the confusion A., Agustian, Y., Putra, A. C. A. Y., and Daffa, F., matrix shows that MobileNetV2 not only achieves “Vegetation classification algorithm using convolutional higher accuracy but also performs better precision, neural network ResNet50 for vegetation mapping in Bandung district area”, Jurnal Infotel, 2022, Vol.14, No.2, recall, and F1 score, providing important guidance doi: https://doi.org/10.20895/infotel.v14i2.756 in model selection and hyperparameter settings for Bai, Z.; Zhu, R.; He, D.; Wang, S.; Huang, Z., “Adulteration similar image classification tasks in the future. Detection of Pork in Mutton Using Smart Phone with the CBAM-Invert-ResNet and Multiple Parts Feature VI. CONCLUSION AND RECOMMENDATION Fusion”. Foods 2023, 12, 3594. https://doi.org/10.3390/ foods12193594 Based on the experiments and analysis carried Simbolon, H. F. S., “Classification bootstrap ResNet50 for out in this research, it can be concluded that potato diseases datasets”, Jurnal Mantik, vol. 7, no. 4, pp. MobileNetV2 consistently outperforms ResNet50 in 3334-3341, 2024. doi: Dong, K., Zhou, C., Ruan, Y., and Li, Y., “MobileNetV2 performance on rice, corn, cassava and sugarcane model for image classification”, Information Technology image datasets. MobileNetV2 shows higher and Computer Application, vol. 2, pp. 476-480, 2020. doi: accuracy, precision, recall, and F1 score than https://doi.org/10.1109/ITCA52113.2020.00106 ResNet50, especially at larger image sizes. The size Prottasha, S. I., Hossain, A. B. M. K., Rahman, Z., Reza S. M. S., and Hossain, D. A., “Identification of Various Rice of 448x448 with 50 epochs produces the highest Plant Diseases Using Optimized Convolutional Neural accuracy of 87.23%, while ResNet50 achieves the Network”, International Journal of Computing and best results at 150 epochs with an image size of Digital Systems, vol. 12, no. 1, pp. 1539-1551, 2022. 75x75. doi:https://dx.doi.org/10.12785/ijcds/1201124 Raiaan, M. A. K., Sakib, S., Fahad, N. M., Mamun, A. A., For recommendations, it is recommended that Rahman, A., Shatabda, S., and Mukta, S. H., “A MobileNetV2 be used for image classification with systematic review of hyperparameter optimization a larger image size and an optimal number of epochs techniques in Convolutional Neural Networks”, Elsevier, around 50 to achieve the best performance. Further vol. 11, pp. 1-32, 2024. doi: https://doi.org/10.1016/j.dajour.2024.100470 research needs to be done to explore combinations Mantravadi, B. S., Vintha, V. V. S. A., Logisa, S., Doki, R., of hyperparameters such as learning rate and batch and Mummana, P., “A Robust Approach to Medicinal size. Additionally, data augmentation can increase Plant Classification through the ResNet50 Model”, the amount of training data and help the model Journal Of Technology, vol. 14, no. 4, pp. 27-31, doi: Sreedhar, P. S. S., and Nandhagopal, N., “Classification generalize better. For further validation, testing on Similiarity Network Model for Image Fusion Using  ResNet50 and GoogLeNet”, Tech Science Press, 2022. Indraswari, R., Rokhana, R., and Herulambang, W.,, doi: 10.32604/iasc.2022.020918. “Melanoma image classification based on MobileNetV2 Faisal, K. M., Hossien, E., Islam, M. T., “Identification and network”, Elsevier, pp. 198-207, 2022. doi: Classification of Rice Disease Using ResNet 50”, 10.1016/j.procs.2021.12.132 Internasional Journal of Scientific & Engineering Gulzar, Y. Fruit Image Classification Model Based on Research, vol. 13, 2022. MobileNetV2 with Deep Transfer Learning Technique. Alwan D. S., and Naji, M. H., “Rice Disease Classification Sustainability 2023, 15, 1906. https://doi.org/ by Residual Netwrok 50 (RESNET50) and Support 10.3390/su15031906 Vector Machine (SVM) Modeling”, Journal of Kufa for Dou, S.; Wang, L.; Fan, D.; Miao, L.; Yan, J.; He, H. Mathematics and Computer, vol. 10, no. 1, pp. 96-101, “Classification of Citrus Huanglongbing Degree Based on 2023. doi: CBAM-MobileNetV2 and Transfer Learning”. Sensors http://dx.doi.org/10.31642/JoKMC/2018/100114 2023, 23, 5587. https://doi.org/10.3390/ s23125587 Usha Ruby A., George Chellin Chandran, J., Chaithanya, B. Rashid, J., Khan, I., Abbasi, I. A., Saeed, M. R., Saddique, N., Swasthika Jain, T. J., and Patil, R., “Wheat Leaf M., and Abbas, M., “A Hybrid Deep Learning Approach Disease Classification Using Modified ResNet50 to Classify the Plant Leaf Species”, Computers, Materials Convolutional Neural Network Model”, Research Square, & Continua, vol. 76, no. 3, pp. 3897-3920, 2023, doi: 2022. doi: https://doi.org/10.21203/rs.3.rs-2130789/v1 10.32604/cmc.2023.040356 Chai C. Foong, Ghong K. Meng, and Lim L. Tze, Bipin Nair, B. J., Arjun, B., Abhishek, S., Abhinav N. M., “Convolutional Neural Network based Rotten Fruit and Madhavan V., “Classification of Indian Medicinal Detection using ResNet50”, Control and System Flowers using MobileNetV2”, IEEE, pp. 1512-1518, Graduate Research Colloquium, 2021, doi: 2024. 10.1109/ICSGRC53186.2021.9515280 Parjito, Ulum, F., Muludi, K., Abidin, Z., Alma, R. M., and Kumar, R., Chug, A., Singh, A. P., and Singh, D., “Evaluation Permata, “Classification of Ornamental Plants with of Deep learning based ResNet-50 for Plant Disease Convolutional Neural Networks and MobileNetV2 Classification with Stability Analysis”, International Approach”, International Seminar on Machine Learning, Conference on Intelligent Computing and Control Optimization, and Data Science, 2022, doi: : Systems, 2022, doi: 10.1109/ISMODE56940.2022.10180988 10.1109/ICICCS53718.2022.9788207 Hoang-Tu Vo, Kheo Chau Mui, Nhon Nguyen Thien, Phuc Pham Tien, and Huan Lam Le, “Optimizing Grape Leaf Rachburee, N., and Punlumjeak, W., “Lotus species Disease Identification Through Transfer Learning and classification using transfer learning based on VGG16, Hyperparameter Tuning”, Internasional Journal of ResNet152V2, and MobileNetV2”, IAES Internasional Advanced Computer Science and Applications, 2024, vol. Journal of Artificial Intelligence, vol. 11, no. 4, pp. 1344- 15, no. 2, pp. 922-931 1352, doi: 2022. 10.11591/ijai.v11.i4.pp1344-1352. Hamid, Y., Wani, S., Sommro, A. B., Alwan, A. A., Gulzar, Shahi, T. B., Sitaula, C., Neupane, A., and Guo, W., “Fruit Y., “Smart Seed Classification System based on classification using attention-based MobileNetV2 for MobileNetV2 Architecture”, International Conference on industrial applications”, PloS ONE, vol. 12, no. 2, 2022. Computing and Information Technology, 2022, doi: Doi: https://doi.org/10.1371/journal.pone.0264586 10.1109/ICCIT52419.2022.9711662