Examination of Lemon Bruising Using Different CNN-Based Classifiers and Local Spectral-Spatial Hyperspectral Imaging

Abstract

The presence of bruises on fruits often indicates cell damage, which can lead to a decrease in the ability of the peel to keep oxygen away from the fruits, and as a result, oxygen breaks down cell walls and membranes damaging fruit content. When chemicals in the fruit are oxidized by enzymes such as polyphenol oxidase, the chemical reaction produces an undesirable and apparent brown color effect, among others. Early detection of bruising prevents low-quality fruit from entering the consumer market. Hereupon, the present paper aims at early identification of bruised lemon fruits using 3D-convolutional neural networks (3D-CNN) via a local spectral-spatial hyperspectral imaging technique, which takes into account adjacent image pixel information in both the frequency (wavelength) and spatial domains of a 3D-tensor hyperspectral image of input lemon fruits. A total of 70 sound lemons were picked up from orchards. First, all fruits were labeled and the hyperspectral images (wavelength range 400–1100 nm) were captured as belonging to the healthy (unbruised) class (class label 0). Next, bruising was applied to each lemon by freefall. Then, the hyperspectral images of all bruised samples were captured in a time gap of 8 (class label 1) and 16 h (class label 2) after bruising was induced, thus resulting in a 3-class ternary classification problem. Four well-known 3D-CNN model namely ResNet, ShuffleNet, DenseNet, and MobileNet were used to classify bruised lemons in Python. Results revealed that the highest classification accuracy (90.47%) was obtained by the ResNet model, followed by DenseNet (85.71%), ShuffleNet (80.95%) and MobileNet (73.80%); all over the test set. ResNet model had larger parameter sizes, but it was proven to be trained faster than other models with fewer number of free parameters. ShuffleNet and MobileNet were easier to train and they needed less storage, but they could not achieve a classification error as low as the other two counterparts.

Comparison of 2D and 3D convolutional neural networks in hyperspectral image analysis of fruits applied to orange bruise detection

“…Pourdarbani et al. (2023) dealt with the detection of bruising in lemons on HSIs applying CNN‐DenseNet, resulting in an 85.7% correct classification. Finally, Munera et al.…”

“…In recent years, there has been a surge of research investigating mechanical damage caused by transport, both in real‐world settings and through simulated experiments. For example, studies have explored the effects of transport on watermelons (Sadrnia et al., 2008), kiwi (Lü & Tang, 2012), tomato (Pathare & Al‐Dairi, 2021), apple (Nturambirwe et al., 2023), and lemon (Pourdarbani et al., 2023). Both destructive and nondestructive methods can be used to differentiate between different types of bruises on fruit.…”

Optimization of pneumonia CT classification model using RepVGG and spatial attention features

“…Compared with the VGG series network, ResNet has fewer parameters and achieves better performance. Sheetal et al proposed a new architecture using resnet50 to classify covid-19 X-ray images into three categories, with a classification accuracy of 0.979 ( 19 , 20 ). Serte et al used the ResNet-50 model to predict each CT image with greater robustness and accuracy ( 21 ).…”

“…It is important to note that the 2019 coronavirus disease (COVID- 19), in addition to symptoms like fever, dry cough, and fatigue, has the potential to result in respiratory failure and life-threatening infections in extreme circumstances, making it a worldwide health concern (3). Chest CT is a non-invasive, fast, and highly sensitive method for detecting COVID-19, which can expedite the diagnosis process (4).…”

Effect of the Light Environment on Image-Based SPAD Value Prediction of Radish Leaves

“…HSV and HSL color systems, which are obtained by converting from the RGB color system, are also used for semantic segmentation [12], such as georeferencing and crop quality monitoring [13]. There are also studies using multispectral cameras [14,15] and hyperspectral cameras [16,17], but these methods require high equipment costs.…”