Computer Intelligence-Based Fruit Grading: A Review

Numerous real-world situations have made use of fruit identification systems, including checkout lines at retail establishments, where they can replace hand scanner tags. It can also be used as a support system for the blind. Fruit identification in supermarkets is laborious, although the cashier must define the type of each item to determine its price. The best solution to this issue is a fruit and vegetable recognition system that automates the labelling and pricing calculations. Examples of several image-processing techniques employed by researchers working on autonomous fruit identification are presented in this paragraph.

A model using morphological features, such as area, major and minor axes, mean red and standard deviation, mean green and standard deviation, mean blue and standard deviation, mean hue and standard deviation, mean saturation and standard deviation, mean intensity, and mean saturation, was proposed by Mustafa et al. [4]. This recipe contained five fruit varieties: apples, bananas, mangoes, carrots, and oranges. Samples were collected using a digital camera. Fruits are classified using the Probabilistic Neural Network (PNN) classifiers are based on morphological and visual traits. The accuracy of the investigation was 90% with a seven-fold cross-validation. Based on the Fitness-Scaled Chaotic Artificial Bee Colony (FSCABC) algorithm and Feed-forward Neural Network (FNN), Zhang et al. [5] proposed a hybrid technique for classifying fruits. In total, 1653 images of 18 different fruit species were captured using a digital camera. The 256X256 pixel scaling sped up the process, but decreased the image quality. The images were captured in a rectangular window. A total of 64 color features, 8 form features, and 7 texture characteristics were extracted for each 256X256 pixel image. To preserve 95% of the original features, the feature dimension was reduced using Principal Component Analysis (PCA). A stratified five-fold cross-validation method was employed to enhance the capacity of the FNN to produce data. The accuracy of the proposed approach (89.1%) was higher than that of several other classifiers according to data. Lu et al. [6] presented a fruit categorization tool with the primary goal of quickly and precisely identifying fruits. The primary objective of this project was to categorize fruits using computer vision and artificial intelligence. To lower the misclassification rate, a single-hidden layer feed-forward neural network (SLFN) is proposed against the proposed classifier approach. Offline learning and prediction are the two phases of the proposed system. Images were obtained from the database and processed offline using PCA to preprocess, extract features, and reduce feature dimensions. The classifier received these properties, which were subsequently evaluated. The accuracy of the proposed methodology was 89.5%. When developing an object recognition control system, Khaing et al. [7] took convolutional neural networks (CNN) into account. This article explains the basic methodology for utilizing CNN in the control framework of the characterization process and the simulation findings. The system setup and validation used 971 images of the most well-liked fruit. These methods can be used to classify 30 different fruit varieties. There were approximately 32 different images for each fruit class. There were eight layers in the proposed CNN model. The loads and predispositions of each stratum were assessed according to a random sample. The maximum number of epochs was set to 1000; maximum learning rate, maximum learning decay rate, momentum, maximum weight decay rate, and early stopping patience were all set to 0, 0.005, 0.0001 and 50, respectively. They employed a predetermined learning arrangement in this instance. Because different fruit varieties are similar and external environmental changes, such as illumination, can affect their surroundings, automatic fruit recognition by machine vision is thought to be a difficult issue. Hussain et al. [8] proposed a novel Deep Convolution Neural Network-based fruit recognition system. They used 44406 images from 15 distinct categories as direct input to the DCNN for training and recognition in their experiment, without previously eliminating any features. The DCNN also acquired knowledge of the images' greatest qualities through an adaptation procedure. The proposed method automatically recognizes fruits with a high level of 99% accuracy. Al Haque et al. [9] proposed a CNN-based model to classify five different banana varieties, including cavendish, ladyfinger, sabri, green, and red bananas, and to identify which bananas are the most decaying. The second of the authors' two Deep Learning-based CNN models performed well, with a classification accuracy of 93.4±0.8% and rotten banana recognition accuracy of 98.3±0.8%. Bongulwar et al. [10] proposed a Model for the identification and classification of fruits using the concept of deep learning. The objective is to build an automatic system for feature extraction using convolutional neural networks. The proposed system uses high quality ‘ImageNet’ dataset. The dataset consists of five different categories of fruit images such as apple, banana, grape, litchi and mango etc. comprising of 4760 numbers of images. The dataset is divided into training and validation datasets in which 90% of the images are trained and 10% are validated. The model uses Convolutional Neural Networks to identify fruits from images. The accuracy obtained is 92.23%. Deep learning outperforms machine learning algorithms. To recognize bananas, Vijayalakshmi et al. [11] proposed a five-layer CNN with convolution, pooling, and fully connected layers. A CNN was used to extract features from fruits, including apples, strawberries, oranges, mangoes, and bananas. Various classification methods, such as Random Forest (RF) and K-nearest neighbor (KNN), have been used to identify fruits (KNN). The deep feature RF combination algorithms outperformed the existing systems by 96.98% when our Deep Learning-based RF and CNN methods were compared. Finally, a technique for configuring the advancement and control frameworks for vision-based automated basic leadership frameworks was developed (see Table 1 for more details).

Table 1. Various approaches used for fruit identification

The fruit business has embraced an image processing approach. This is a quick, reliable, and objective inspection method. In recent years, fruit grading has become increasingly dependent on image processing. Grading entails categorizing fruits while considering disease severity, flaws, and contamination on the produce. Grading is a crucial phase of the postharvest procedure. However, manual fruit grading is a laborious and inaccurate process. Therefore, it is necessary to alter the automated speedier system. An automatic image-processing system is a dependable method for fruit sorting and grading. This section examines how image processing has evolved in the agricultural and food-processing sectors.

Razak et al. [20] used fuzzy analysis to propose autonomous mango grading. Color and skin features were collected using this scale. The area of the sample image was used to calculate mango size. Subsequently, the RGB component of the image was removed, and the average of the three-color components was determined. The edge detection approach was used for shape analysis. Mango was graded into different classes using fuzzy inference methods, which provided an overall accuracy of 80%. Zheng et al. [21] developed a mango rating system. The L*a*b* color model and fractal dimension were employed for grading. The fractal dimension and color accuracy obtained using SVM were 85.19% and 88.89%, respectively. Pandey et al. [22] proposed an automated mango grading technique that uses image processing. By applying the CIE Lab color model to extract color features, the Dominant Density Range approach divides mangoes into Healthy and Diseased groups. The color information of an object is efficiently represented using the L*a*b* color space. The color ratio determines mango health and sickness. After the discovery of healthy mangos, size-based grading was assessed using area and diameter. According to an experimental finding, the Dominant Density Range approach and the CIE Lab color model can be used to grade mangoes successfully. Mangoes can be classified into target classes with an average accuracy of 92.37% using the suggested approach. Nandi et al. [23] suggested an automatic prototype system for grading and sorting mangoes that makes use of fuzzy logic. Using a CCD camera set on top of a conveyor belt that is moving mangoes, the automated system collects video images. It then processes the images to gather pertinent attributes sensitive to mango size and maturity level. The Gaussian Mixture Model (GMM) was used to estimate the parameters of the various classes to forecast maturity. Mango fruits are automatically sorted and graded using fuzzy logic techniques, and the mango size is estimated from the binary image of the fruit. A new automated approach for sorting and grading mangoes based on computer vision algorithms is provided by Pauly and Sankar [24]. This method can be used to replace India's current manual sorting and grading process. Alphonso mango, a premium type of mango shipped from India, is the target market for the system. With an accuracy of 83.3%, the created system could sort Alphonso mangoes and recognize defective skin up to a minimum area of 6.09384510 -4 sqcm. Agilandeeswari et al. [25] suggested an autonomous multi-class support vector machine (SVM) based mango grading system. The mangoes were classified into three groups using a multi-class SVM classifier —very good, good, and bad—after pre-processing, segmentation, feature extraction, and classification. The proposed methodology has a 97% accuracy rate. Panda and Sethy [26] developed a system for Carica papaya grading using the Artificial Bee Colony algorithm (ABC) to classify papaya fruits from digital images. Our preliminary analysis of the image features suggests that the ABC algorithm's criteria for sorting papaya fruits into different grades could include impacted area, shape, and texture. Fuzzy logic, naive Bayes classifier, and support vector machine (SVM) are evaluated for their effectiveness in rating papayas. The input papaya was divided into two groups during the classification process: healthy and defective. The SVM classifier provides an accuracy of 93.5% in all grading steps, the naive Bayes classifier provides an accuracy of 92%, and fuzzy logic provides an accuracy of 86.04%. The accuracy of the proposed optimization techniques for various papaya fruit picture databases is 94.04%. Naik [27] proposed two techniques for mango fruit grading. The CNN was initially trained using mango samples divided into classes I, II, III, and IV. In comparison, the second method used three stages for mango grading. The CNN is trained using labels for the shape parameter in the first phase, which are both well-formed and deformed; labels for the size parameter in the second phase, which are small, medium, and large; and labels for the maturity parameter in the third phase, which are ripe, partially ripe, and unripe. These three processes are used to base the grade decision. Ucat and Dela Cruz [28] presented a post-harvest grading categorization of Cavendish bananas using deep learning and TensorFlow. Python OpenCV and TensorFlow were employed to create an algorithm and increase classification accuracy. The number of bananas utilized was 1116, and there were 279 images in each of the three categories (cluster Class A (part of hand), class A big-hand or small-hand, and Class B big-hand or small-hand). Four procedures are used for the sample images: picture thresholding, feature extraction, classification prediction, and testing. Regarding the expected accuracy across all classes, the extraction of the finger size value (small or large hand) was more significant than the extraction of the surface defect value (A and B). In all four classes across all trained validation and test data, the final classification displayed accuracy above 90%. Based on maturity, size, form, and flaws, Supekar and Wakode [29] presented a novel approach to evaluate Dashehari mangoes. 85 Dashehari mangoes were compiled into a dataset. Mangoes were viewed from two angles, making the devised technique more dependable. Before determining the mango grade, each grading parameter category for mangoes was established (grades 1, 2, 3, and 4). The random forest classifier accomplished perfect categorization based on ripeness and form. However, several medium and small mangoes have been incorrectly classified based on their size. The accuracy of the test set classification based on ripeness, size, and form was 100%, 98.19%, and 99.20%, respectively. The segmentation method based on K-means clustering also yielded good results in locating mango flaws. The mango grade was determined using a formula for mango grading. Sixteen of the 18 test mangoes were correctly assessed, resulting in a grading accuracy of 88.88%. Therefore, mango grading, image processing, and machine-learning approaches have been successfully used. This technique can be used for other mango types in future studies. Firmness and sweetness are the two other grading standards that can be used. Creating a multi-variant grading system that determines the mango variety before carrying out the proper grading is possible. Four groups were created by Raghavendra et al. [30] to categorize mangoes according to their level of ripeness. The work was tested by conducting lengthy experiments on a recently developed group of 981 images of the Alphonso mango variety, categorized into four categories: under-ripe, perfectly ripe, overripe with internal issues, and overripe without internal abnormalities. Mangoes were divided into four classes using a hierarchical technique. At each level of classification, the characteristics from the L*a*b color space were extracted. Various classifiers and their potential combinations were tested for classification at each level. The study discovered that when categorizing mangoes into under-ripe, perfectly ripe, and over-ripe, the thresholding classifier outperforms the Support Vector Machine (SVM) classifier in classifying over- ripe with internal faults and over-ripe without internal defects. Furthermore, a traditional single-shot multiclass classification strategy using SVM was tested to highlight the advantages of the hierarchical approach. The results of the experiment revealed that several modern models and the matching conventional single-shot multi-class classification strategy outperformed the hierarchical method, with an accuracy of 88%. Mesa and Chiang [31] suggested a non-invasive automated banana grading system that uses deep learning methods, RGB, and hyper spectral imaging. In accordance with international rules, a real dataset of pre-classified banana tiers (Class 1 for export-quality bananas, Class 2 for the local market, and Class 3 for faulty fruits) was used. Fewer samples were used with the multi-input model than with earlier methods in the literature, and the results showed an excellent overall accuracy of 98.45%. Gururaj et al. [32] provided a system of grading mangoes that is intelligent and can determine the overall quality of the fruit based on its appearance, such as maturity ripening stage, shape, color, texture, and fault features, among others. The efficiency of the proposed system was improved by the characteristics of the CNN, which had 93.23% accuracy for variety recognition, and 95.11% accuracy for quality grading. The automatic mango sorting and grading model that Iqbal and Hakim [33] reported using a DL technique considered eight categories of harvested mango qualities, such as shape, size, color, and texture. Image rotation, translation, zooming, sharing, and horizontal flipping were accomplished using data augmentation techniques. The Inception v3 CNN architecture achieved 99.2% sorting accuracy and an average grading accuracy of 96.7% when compared to the VGG16, ResNet152, and Inception v3 techniques while using supplemented data. A low-cost machine vision system based on deep learning was proposed by Ismail and Malik [34] to grade fruits according to their external appearance or freshness. The system was trained and assessed (apples and bananas) on two datasets. Using the EfficientNet model, the average accuracies for the test sets for apples and bananas were 99.2% and 98.6%, respectively. While using stacking ensemble deep learning approaches, they also found a minor increase in the recognition rate of 0.03% for apples and 0.06% for bananas. It was discovered that the new approach outperformed earlier methods used on identical datasets. Furthermore, real-time testing of actual samples revealed that the accuracy was 96.7% for apples and 93.8% for bananas, demonstrating the effectiveness of the proposed system, as displayed in Table 3.

Table 3. Various techniques applied for grading fruits

Applying image processing methods will ultimately improve the qualitative and quantitative evaluation of fruit and vegetable recognition, categorization, and disease detection in fruits and vegetables among local goods in agricultural regions. The advantages of this vision-based technology include speed, persistence, and non-destructiveness. These obstacles include poor lighting and a variety of capture sites, whereas variability, cropping, and osculation do not affect the ability to distinguish various types of fruits. The survey found that most researchers who study fruit inspection and/or analysis follow the same methodology, as shown in Figure 5. There are some research challenges and limitations that can improve the current state-of-the-art.

5.png

Figure 5. A generally acceptable processes for fruit recognition

6.1 Challenges

• Reduced effort, labor costs, and other expenses are benefits of employing computer vision in fruits and vegetables. The conventional method requires sufficient time to process defects, whereas computer vision automatically reduces the processing time for defect identification.

• Fruit and vegetable databases are more challenging for classification and identification because of their unique size, shape, color, and defects.

• Flaw detection can be achieved using various datasets of fruit components. For example, a training model can be built using a width multiplier, and checking can be performed at a faster rate.

6.2 Limitations

• Existing computer vision techniques require the creation of a dataset; however, they consume more time. This is one of the main disadvantages of the proposed method.

• Another disadvantage is that if the processing time to process an image is longer, it may lead to a vacillation effect.

• Fruits are more environment-dependent, which is a significant drawback because the same computer vision technique may result in variable degrees of accuracy in the dataset.

• Sometimes, the accuracy can be improved by improper exploration of the internal fruit structures, which increases the likelihood of finding flaws and performing quality analysis.

• It is necessary to be more trustworthy and efficient in creating new algorithms and techniques for data extraction, processing, and analysis.

• The present feature extraction methods cannot extract spectral data from different regions of the fruit. Therefore, it is necessary to create a multichannel spectroscopic system that can instantly check for internal faults in various fruits.

• To further develop this (neural network) approach as a technical framework that might be used for a variety of industrial applications, such as the identification of fruit and vegetable quality deficiencies.