Fig 1.
Overall block diagram of our proposed method for fruits image classification.
Note that GAP stands for Global Average Pooling to convert the 3D features to 1D features and Conv. layer denote the last convolution layer of MobileNetV2 used in our work. Note that the fruit image used here is Republished from [44] under CC BY license, with permission from [Georg Waltner], original copyright[2022].
Table 1.
The complete network structure of MobileNetV2.
Fig 2.
Pre-trained MobilenetV2-architecture used as convolutional module in this work [15].
Please note that nth bottleneck residual block is denoted as ‘BR-Block-N’ in Fig 2. Note that the fruit image used here is Republished from [44] under CC BY license, with permission from [Georg Waltner], original copyright[2022].
Fig 3.
Model training accuracy and loss per epoch of our model on first set of D1.
Fig 4.
Sample fruits images abstracted from Dataset D1.
Note that (a), (b), and (c) denote “potato”, “diamond peach” and “watermelon” fruits classes, respectively.
Fig 5.
Sample fruits images abstracted from Dataset 2 (D2).
Note that (a), (b), and (c) denote “pomegranate good”, “pomegranate bad” and “guava good” fruits classes, respectively. Republished from [51] under CC BY license.
Fig 6.
Sample fruits images abstracted from Dataset 3 (D3).
Note that (a), (b), and (c) denote “Avacado”, “Banana” and “Grape” fruits classes, respectively. Note that the fruit image used here is Republished from [44] under CC BY license, with permission from [Georg Waltner], original copyright[2022].
Table 2.
Detailed hyper-parameters used in our study.
Table 3.
Precision, Recall, Kappa-score, MAF1, WAF1, and Acc. on D1, D2, and D3 using averaged metrics over five runs (%).
Table 4.
Model parameters (’000) and running time (seconds) for each model.
Table 5.
Reported accuracy of state of the art methods using classification accuracy (%).
Table 6.
Precision, Recall, Kappa-score, MAF1, WAF1, and Acc. on D1, D2, and D3 using averaged metrics over five runs (%).
Table 7.
Ablative study of our proposed model using average performance metrics (precision, recall, MAF1-score and accuracy) on D1.
Table 8.
Averaged class-wise Precision, Recall and F1-score of our model on test samples of D1 using over five runs (%).
Fig 7.
The GradCam [55] visualization examples for Fruits images from D3.
Each row contains the original fruit image, its corresponding heatmaps extracted by Convolution module, and Attention module. Note that the fruit image used here is Republished from [44] under CC BY license, with permission from [Georg Waltner], original copyright[2022].