In 2017, global production of chickpea (
Cicer arietinum L.) was estimated in almost 15 million tons/year. In addition, the average yield of chickpea in Iran is lower than the world average. This difference between the mean and potential yield of this product may be due to its susceptibility to biotic and non-biotic constraints [
1]. About 75% of chickpea is used for human consumption and from the remaining, 14% is used for livestock consumption, 7% is used for seed production, and 4% is wasted at different stages of processing [
2]. The importance of chickpea grading is highlighted when consumers go to different distributors and ask them to prepare their order. In other words, different varieties of chickpea have different applications, properties and nutritional values. Therefore, it is important to correctly identify the product variety. This assessment is often made by human experts, but for various reasons, such as expert psychological changes, fatigue, and so on, errors in the evaluation of different varieties may occur. For this reason, it might be important to provide rapid detection systems using new technology. One can consider at least two main reasons why it might be important to provide rapid detection systems using new technology: first, different applications and price of each chickpea variety; second, separation of each chickpea variety when stored in the same tank, once they have been harvested. Different methods have been proposed for the grading of different products. Some of these systems are based on physical and mechanical properties [
3,
4,
5]. Although these systems are simple and inexpensive, they are not popular because of their often low efficiency and capacity. For instance, in order to separate ripe fruit from immature, fruits were fed in a row on a v-shaped conveyor. At the sorting points, the piston behind the belt vibrated at a predetermined resonance frequency, producing movement over the fruits. Vibrant immature fruits were cut into the lateral conveyor belts and separated from ripened ones. Another example was the use of elasticity feature in the sorting of ripe fruit. Fruits were given a free fall onto a rotating cylinder. As the rotary cylinder moved, the fruits had a specific downward parabola motion. Ripe fruit with soft texture would be directed to closer outlet and unripe fruit with firm texture would be directed farther to the second outlet. Since these classical food separation systems have a high response time as compared to new computer vision techniques, thus classical system capacity is low. In addition, it is also possible to damage fruit piece texture due to free fall or mechanical vibration [
6]. In recent years, the application of new technologies such as computer vision and machine learning has been increasing in various areas of the agricultural and food industries. These systems play an important role in the precise and uniform control of food quality by eliminating manual inspection and providing an appropriate objective, automatic, and non-intrusive way to distinguish the desired product characteristics (color, size, shape, texture, and pests) for separation and classification [
7,
8] purposes. Iannace et al. developed a model based on the artificial neural network to detect unbalanced blades in an unmanned aerial vehicle (UAV) propeller. An air collision or a defect in the propellers of a drone can cause the aircraft to fall to the ground, and thus its consequent destruction [
9]. LeCun et al. recognized document content using a gradient-based learning rule [
10]. Uijlings et al. introduce a selective search which combines the strength of both an exhaustive search and segmentation for object recognition that results in a small set of data-driven, class-independent, high quality locations, yielding a 99% recall value [
11]. Kurtulmus and Unal [
12] studied on seven different varieties of rapeseed. Mass level imaging method was selected instead of kernel surface imaging for rapeseed classification. Mass surfaces were constituted by spreading rapeseed samples in a 15–20 mm layer in a square frame in order to prevent the scanner light from passing through the mass layer. Before scanning of each individual sample, the scanner glass was thoroughly cleaned. An imaging chamber was designed using a charged-coupled device (CCD) scanner and a personal computer under laboratory-controlled conditions. Samples were placed in a frame of 36 × 36 mm
2 to obtain a uniform image. A total of 525 samples were obtained by scanning 75 rapeseed samples from each variety in lab. Three classes of texture properties were extracted from rapeseed image samples: Gray Level Co-ocurrence Matrix (GLCM), defined as a space frequency matrix between adjacent pixels in a digitized image, Gray Level Run Length Matrix (GLRLM), defined as the number of runs with pixels of gray level
i and run length
j at a given image direction, and local binary patterns, that transform an image into an array of integer labels describing small-scale appearance on an image. Next, three classifiers including support vector machine classifications, k-nearest neighbor, and stochastic gradient descent, were used to classify seven rapeseed varieties. Results showed that the classifiers clustered samples with good accuracy. Other researchers have proposed a new approach based on image analysis techniques to discriminate defective lentil seeds from healthy ones. They extracted several color, shape and texture features. These features were then used as inputs for Support Vector Machine (SVM) classification. Results showed that their method achieved accuracy was 98.9% [
13].
Liu et al. [
14] identified different varieties of soybean as well as healthy versus unhealthy seeds using image processing. For this purpose, 857 soybeans were collected, and then imaging was performed. Shape, texture, and color characteristics were extracted. Finally, classification was performed using an artificial neural network classifier with a 97% accuracy. Golpour et al. [
15] used texture features of digital images to identify paddy, brown, and white rice varieties. In this study, 2.5 kg with 11–12% moisture content of each variety was prepared. A total of 90 3D images of each variety were captured under standard conditions. A total of 1350 images of 540 × 390 pixels were generated. The classification was performed with a 97.67% accuracy.
The aim of the present paper is to develop a computer vision system in order to classify three similar varieties of chickpea (Adel, Azad, and Arman) using an image processing and machine learning ensemble ANN majority voting (MV) approach between hybrid artificial neural network- particle swarm optimization (ANN-PSO), hybrid artificial neural network-ant colony optimization (ANN-ACO) and hybrid artificial neural network-harmonic search (ANN-HS) outputs.