Detection of Almond Varieties Using Impact Acoustics and Artificial Neural Networks

Transcription

1 International Journal of Agriculture and Crop Sciences. Available online at IJACS/213/6-14/ ISSN X 213 IJACS Journal Detection of Almond Varieties Using Impact Acoustics and Artificial Neural Networks Ali Reshadsedghi 1, Asghar Mahmoudi 1 1. Department of Agricultural Machinery, Faculty of Agriculture, University of Tabriz, Tabriz, Iran Corresponding author a_mahmoudi@tabrizu.ac.ir, sedghi_al@yahoo.com ABSTRACT: This study was conducted to detect and classify different almond varieties based on their shell thickness and hardness by using an impact acoustic system combined with artificial neural networks (ANNs). To develop the ANN models, a total of 6 almond sound signals, 1 samples for each genotype, was recorded. Almond nuts are dropped onto a steel impact plate through a pipe. The impact acoustic signal generated by the system is captured by a microphone and processed by a PC. Features such as amplitude, phase and power spectral density of almond nut varieties were extracted from analysis of sound signal in both time and frequency domains by means of Fast Fourier Transform. Principal component analysis method was used to reduction of features as input vector to ANN models. The optimal model of artificial neural networks was selected after several evaluations based on minimizing of mean square error (MSE), correct classification rate (CCR) and coefficient of correlation. The topology of a Multilayer perceptrons neural network exhibited highest accuracy (r-values) and least error (MSE) on cross validation data set. The correct classification rate for hard, semi soft and soft shell almond nuts were 93.78%, 94.51% and 99.63%, respectively. To confirm the system potentiality for distinction and classification of almond varieties in three classes: Hard, Semi-soft and, two sets of experiments (with 3 and 6 almond genotypes) were conducted. Results indicated that by increasing almond varieties number for classification, the system accuracy is going to be reduced. Keywords: Almond; Artificial neural networks; Detection; Impact Acoustics INTRUDUCTION The almond, (Prunus amygdalus), is a species of tree, native to the Middle East and South Asia. The fruit of the almond is a drupe, consisting of an outer hull and a hard shell with the seed inside. Shelling almond refers to remove the shell to reveal the seed. Almonds are sold shelled, or unshelled. Iran is one of the most important regions for origin and diversification of wild almond species in the world. Over 2 species, naturally distributed in many regions, have been identified to date in Iran (Madam et al., 211). According to the FAO (211), Iran with an annual production of about 168, MT, is the third producer of almond in the world after United States of America and Spain, and Iranian almonds have the highest value ($686 /ton) in the world, but their portion in world market is only.22 percent. This problem may be due to the lack of a suitable sorting and packing system for almond in Iran. Several varieties of almond can be found in the traditional and modern gardens of Iran. In the harvest stage and after picking the almonds from the trees or in the silo stage, it may that these varieties be mixed together. In the almond s marketing and processing industry as other agricultural products, it is necessary that products be uniform in the case of variety. Because of their different characteristics such as shell hardness and thickness, shape, size, kernel percentage and kernel taste, influence on their marketing price so that almonds with softer shell have more price. Shell hardness of almond nut changes in relation to almond cultivars and cracking operation is affected by geometric and mechanical properties of almond cultivar (Altuntas et al., 21). So, it seems that for operating a nut cracking and grinding system, it's primarily necessary to sort the different varieties of almond based on their shell hardness and thickness. According to US standards (1997) for grading almonds in the shell, "Similar varietal characteristics" means that the almonds are similar in shape, and are reasonably uniform in degree of hardness of the shells, and that bitter almonds are not mixed with sweet almonds. For example, hard-shelled varieties, semi-soft shelled varieties, soft-shelled varieties and paper-shelled varieties are not mixed together, nor are any two of these types mixed under this definition. Inshell almonds are classified into two types according to the hardness of the shell, as /semi-soft inshell almond that can be easily cracked with the fingers or with a

2 Intl J Agri Crop Sci. Vol., 5 (),, 213 nutcracker and Hard inshell almond that can be cracked only with a hammer or similar devices. Various techniques including optical, mechanical, electrical and acoustical are found increasingly useful in the food industry, especially for applications in quality inspection, damage detection and sorting. Researches in this area indicate the feasibility of using such systems to improve product quality while freeing people from the traditional manual inspection of agricultural materials (Omid et al. 21). The objective of this study was detection and classification of different almond varieties based on their shell thickness and hardness by a low cost, high speed and an accurate intelligent system. Recently, acoustical experiments have been increasingly employed in agriculture. Pearson (21) developed an acoustic sorting system to separate pistachio nuts with open shells from those with closed shells. The system included a microphone, digital signal processing hardware, material handling equipment, and an air reject mechanism. Linear discriminant analysis was used to classify nuts using three features extracted from microphone signal during the first 1.4 ms after impact nut with a steel plate. The classification accuracy of system was approximately 97%. Afterwards, an algorithm using speech recognition technology to distinguish pistachio nuts with shells from those with open shells was developed by Cetin et al. (24). Features extracted from the sound signals consisted of mel-cepstrum coefficients and eigen values obtained from the principle component analysis (PCA) of the autocorrelation matrix of the sound signals. The classification accuracy of closed shell nuts was more than 99% on the validation set, which did not include the training set. The same impact acoustics based system was later extended to separate cracked hazelnuts shells from undamaged ones (Kalkan & Yardimci, 26), underdeveloped ones from full hazelnuts (Onaran et al., 26) and wheat inspection for detection of IDK (insect damaged kernel) from undamaged kernels (Pearson et al., 27). It was observed that the algorithms described by Pearson (21) and Cetin et al. (24) did not produce high classification accuracy in hazelnuts. Although the mechanical structure was similar, the authors reported that the signal features used for pistachio classification did not work well in wheat inspection. The results obtained by these works emphasized the importance of signal processing methods of the impact acoustic signal to achieve higher accuracies in food inspection (Omid et al., 21). An intelligent separation system, based on artificial neural networks (ANNs) was presented by Mahmoudi et al. (26), for classifying four different varieties of Iranian pistachio nuts. Features of pistachio nut varieties were extracted from analysis of sound signal in both time and frequency domains by means of Fast Fourier Transform, power spectral density and principal component analysis methods. Selected optimal ANN for classification was of configuration and net weight average of system accuracy was found to be 97.51%. This intelligent system was applied by Omid et al. (21) for sorting open and closed-shell pistachio nuts. Their system combined acoustic emissions analysis, Principal Component Analysis (PCA) and Multilayer Feed forward Neural Network (MFNN) classifier. The best classifier had a structure. The results indicated the size of pistachio nuts has no effect on the accuracy of the sorter. Ebrahimi & Mollazadeh (21) presented an algorithm based method to sort four varieties of almond (Yalda, France, Shokofeh, and Shahrood 15) using impulse acoustic signals, selected features and rules generated from a decision tree, and fuzzy inference system. Results showed that the best decision tree will be obtained when impact plate was Stainless steel and fall height is 24 cm. In this condition, the fuzzy inference engine performance was found to be encouraging and its accuracy in the classification of almond varieties was 84.16%. Kalkan et al. (28) introduced an adaptive time frequency plane feature selection algorithm obtained from impact acoustic signals to separate damaged/cracked hazelnut kernels from regular ones. The adaptation in time and frequency was achieved by combining local cosine packets and an undecimated wavelet transform. The algorithm achieved a throughput rate of 45 nuts/s and a classification accuracy of 96% with the 3 most discriminative features, a higher rate than those provided with prior methods. Feasibility of laboratory detection of damaged seeds in precision planters caused by malfunction of seed metering device was investigated by Karimi et al. (212) using impact acoustics and artificial neural networks. Correct detection rate of proposed ANNs model ( ) for undamaged and damaged seeds was and 1 respectively. MATERIALS AND METHODS Almond varieties Six almond genotypes in three classes (Hard, Semi- hard/ Semi-soft and ), each class included two genotypes, collected from an orchard of Sahand horticultural researches station of East-Azarbaijan province, Iran, in September 212, were used in this study (shown in Fig. 1). The number of almond nuts, for each variety, used for each test was 1. Some geometric and physical properties of almond nuts used in this study are shown in Table 1. The moisture content of inshell almond samples measured at the test time was about 4-5 percent (dry basis). 19

3 Intl J Agri Crop Sci. Vol., 5 (),, 213 Hard1 Semi soft1 1 Hard2 Semi soft2 2 Figure 1. Six almond varieties in three classes (Hard, Semi soft and ), each class includes two genotypes. Genotype class Table 1. Average values of some geometric characteristics and physical properties of almond nuts. Geometric characteristics Major diameter (mm) Intermediate diameter (mm) Minor diameter (mm) Shell thickness (mm) Physical properties Total nut mass (gr) Kernel to shell weight ratio Hard Hard Semi-soft Semi Bulk density (gr.cm -3 ) Tube Almond PC Feeding platform Impact plate Microphone Acoustic chamber Figure 2. The schematic of experimental impact acoustic system. 11

4 Intl J Agri Crop Sci. Vol., 5 (),, 213 Sensor Data acquisition Signal processing Feature extraction Output Interpretation system Detection system Figure 3. Block diagram of the detection system. The mass of a single almond nut is negligibly compared to the mass of the impact plate. Hence, the possibility of vibrations from the plate interfering with acoustic emissions from nuts was minimized. The microphone (VM-34CY model) was sensitive to frequencies up to 1 khz and to eliminate the environmental noise effects, it was installed inside an isolated chamber which was filled with glass wool. Microphone output signal was sent to a PC (Intel Core(TM) i7-267qm 2.2 GHz, Windows 7 operating system) based data-acquisition where it was digitized at a sampling frequency of 44.1 khz, with 16 bit resolution. A block diagram describing the overall detection system is shown in Figure 3. The emitted sound signals were acquired by the microphone, digitized by the sound card and saved by using Matlab (version 212a) data acquisition toolbox. When almond nuts hit the impact plate, the amplitude of the microphone output signal ranged from to 1. V. Data acquisition began when the microphone output was higher than.1 V. This threshold level was sufficient to trigger acquisition for virtually all nuts, whilst preventing false triggers from ambient sound. Since the maximum frequency of the sound card was 44.1 khz, upon receiving a trigger signal, the computer acquired 15 data points during 34 ms from every sample in the time-domain which was sufficient to extract signal features. Matlab software was used for data collection and management. System description The experimental impact acoustic system (shown in Fig. 2) consists of a feeding platform, an impact plate, an acoustic unit and a PC based data-acquisition. A PVC tube with 5 cm length and 4 cm diameter on an adjustable height base which was inclined 33º above the horizon used as feeding platform. A stainless steel plate with dimensions mm is used as the impact plate. Almond nuts are dropped onto the impact plate through the pipe. The impact acoustic signal generated by the system is captured by a microphone and processed by a PC. The impact plate is fixed to the ground at a 14 º angle. Impact plate and feeder tube inclined angles were determined by trial and error so that each nut would impact one time to the plate. Signal processing and feature extraction For extracting potential features, recorded sound signals were processed and analyzed in both time and frequency domains. Emitted sound signals of 1 almond samples, on average, (for 3 classes: Hard, Semi-soft and ) in time domain are shown in Fig

5 Average FFT Amplitude Power Spectral Density (db/hz) Amplitude (V) Intl J Agri Crop Sci. Vol., 5 (),, Hard Semi soft Time (ms) Figure 4. Emitted sound signals of 1 almond samples, on average, (for 3 classes: Hard, Semi-soft and ) in time domain. Comparing the sound signals indicated that all almond varieties in this study had approximately the same structure of signal spectrum in time domain but peak values of soft almond signal were relatively lower than the two other's. This was maybe due to their lower mass or shell hardness compared with hard and semi-soft ones (Table 1). Also, the signal phase of soft almond was different from hard and semi-soft ones after time 8 ms. Therefore, it seemed that the preliminary attempts to use only time-domain features couldn't be successful. However, in order not to lose any useful transient feature, all 15 data point amplitude values were considered as features. A 124-point Fast Fourier Transform (FFT) was computed from each sound signal. The magnitude, Y(ω), power spectral density (PSD) and phase angle of each spectrum was computed. Fig.5 (a) shows the computed FFT amplitude, corresponding to time-domain signal shown in Fig. 4. Power spectral density was calculated as: ( ) ( ) (1) Where * means complex conjugation. According to symmetry of sound signal in frequency domain, only 512 data from 124 data points were used for calculating PSD and phase angle (Fig. 5(b) and 5(c)) Hard Semi soft Frequency (khz) (a) Frequency (khz) (b) Hard Semi soft 112

6 Number of PCs Phase angle (Radian) Intl J Agri Crop Sci. Vol., 5 (),, Hard Semi soft Frequency (khz) (c) Figure 5. (a) FFT amplitude; (b) Power spectral density; (c) FFT phase angle of sound signals in frequency domain. A total of 2524 features were obtained for each almond. For real time systems, this dimension of the input vector is too large, but the components of the vectors are highly correlated. Principle Component Analysis (PCA) was used to reduce the dimension of the input vectors to a maximum of 5 features. After normalization of data, PCA analysis was performed on data using Matlab software (MathWorks, 212). The PCA results are shown in Fig. 6. Principal components (PCs) for different combinations of signal Amplitude, PSD and Phase angle were considered. From Fig. 6, by using PCA analysis, for example, one can express 98% of total variations in the input data set with only 7 amplitude components instead of 15 data points, one PSD and 4 Phase angle components instead of 512 data points of each one. In this case, only.2% of variances between features would be discarded. To find the best combinations and the minimum number of PCs with highest accuracy, various combinations of PCs were used as feature vector in a multilayer feed forward neural network (MFNN) classifier, off-line PSD Amplitude Phase Angle Percentage of eliminated components variance Figure 6. Relation among number of selected PCs and components variances. Back propagation neural networks The multilayer perceptron (MLP) is one of the most widely implemented neural network topologies (Haykin, 1999). MLPs are normally trained with the back propagation algorithm. In fact the renewed interest in ANNs was in part triggered by the emergence of back propagation learning rule. The back propagation rule propagates the errors through the network and allows adaptation of the hidden processing elements (PEs). Two important characteristics of MLP are (i) use of nonlinear PEs such as logistic or hyperbolic tangent and (ii) their massive interconnectivity, i.e., any element of a given layer feeds all the elements of the next layer. The MLP is trained with error correction learning, which means that the desired response for the system must be known a priori. After adequate training, the network weights are adapted and employed for cross validation in order to determine the ANN model overall performance. Gradient descent with momentum (GDM) learning rule is an improvement to the straight GD rule in the sense that a momentum term is used to speed up learning and stabilizing convergence (Rumelhart et al., 1986). Therefore, the GDM method of learning is used throughout this study. The topology of MLP neural 113

7 Average of Min MSEs Average of Min MSEs Intl J Agri Crop Sci. Vol., 6 (14), , 213 network was included a three layer network incorporating a single hidden layer of processing elements. Each PE has a weighted connection to every PE in the next layer and each performs a summation of its inputs passing the results through a transfer function. In order to minimize ANN training time, only one hidden layer was considered in the network. The best network weights are saved on the parameter variation, run, and epoch when the cross validation error is minimum (Table 3).The number of neuron in hidden layer was determined using an exhaustive search from 2 to 2 nodes at 1 epochs. By using information about mean square error (MSE) of cross validation (CV) for different ANN models, the number of PEs in hidden layer could be selected. To evaluate the potential of experimental acoustic system for distinction and classification of almond varieties, at first only 3 data set related to 3 almond classes based on their shell stiffness, Hard1, semi-soft2 and soft1, (shown in Fig. 1) were tested by ANN models and then experiments were repeated for all six almond varieties data sets (for each almond class, two different genotypes). Considering data obtained in two separate tests by 3 and 6 groups of almond, network with 16 and 12 PE's in hidden layer respectively had the least standard deviation error as well as high stability (shown in Fig. 7 a and b). In developing ANN models, the linear function at the input layer and the nonlinear hyperbolic tangent function at both hidden and output layers were used as transfer functions. Learning rate was.7 throughout the momentum learning rule. As an additional guard against over-fitting, the data sets were divided into three randomly selected data sets; 6% of data were used for training, 15% for cross validation and 25% were used for testing. Neurosolution software was used for design and testing of ANN models. Table 3. Parameters of best training for neural networks (for 3 and 6 almond genotypes data sets). Best Networks Training Cross Validation 3 sets 6 sets 3 sets 6 sets Hidden 1 PEs Run # Epoch # Minimum MSE Final MSE Hidden 1 PEs Hidden 1 PEs Training + 1 Standard Deviation - 1 Standard Deviation Cross Validation + 1 Standard Deviation - 1 Standard Deviation (a) (b) Figure 7. Average of minimum MSEs with standard deviation boundaries: (a) for 3 almond genotypes data sets; (b) for 6 almond genotypes data sets. To find the best combinations of potential features and optimal ANN configuration, many different combinations of principle component features were selected and tested by neural network (Table 4). These features were fed to the ANN models and their performances were compared based on mean square error (MSE), correlation coefficient (r) and correct classification rate (CCR). Expression used to calculate the MSE is given by follow equation: 114

8 Intl J Agri Crop Sci. Vol., 6 (14), , 213 ( ) (2) Where P is the number of output neurons, N is the number of exemplars in data set, and t ij and y ij are the network and target outputs for exemplar i at neuron j, respectively. RESULTS AND DISCUSSION As previous mentioned, to confirm the system potentiality for distinction of almond varieties, two sets of experiments (with 3 and 6 almond genotypes) were conducted to discriminate and classify almonds in three classes: Hard, Semi-soft and. The results of each experiment are as follows separately: Experiment 1 Testing with 3 sets of almond nuts, each of Hard, Semi-soft and classes was carried out first. This experiment was designed to find the optimum Multilayer Neural Network configuration to be used later during online experiments. The summary of results is shown in Table 4. Among the 13 different configurations examined in Table 4, the best PC combinations for the input vector were found as 6 signal PSD and 29 amplitudes features (.5% variance omitted). Therefore the ( )-MLP topology shown in Figure 8, exhibited highest accuracy (r-values) and least error (MSE) on CV data set. The correct classification rate for hard, semi soft and soft shell almond nuts were 93.78%, 94.51% and 99.63%, respectively. a Features a PCs Table 4. Results of selecting various features as input vector to MLNN. b (MSE) (r) Rate (CCR)% Hard Semisofsofsoft Hard Semi- Hard Semi- Mean Square Error Correlation coefficient Correct Classification Psd.1 Amp Ang.1 Psd.1 Amp Ang.1 Psd.1 Amp Ang.5 Psd Amp.1 Psd Amp.1 Ang.1 Psd Amp.5 Ang.5 Psd Amp.1 Psd Amp.1 Ang.1 Psd Amp.5 Psd Amp.5 Ang.1 Psd Amp.5 Ang.5 Psd Amp.1 Ang.1 Psd.1 Amp PSD, Amp and Ang are power spectrum density, amplitude and phase angle respectively. Numbers next to them indicate the percentages of eliminated components variance. b Total number of principal components (x), each feature separately, selected for (x-16-3)-mlp topology. 115

9 Intl J Agri Crop Sci. Vol., 6 (14), , PSD Hard 2 Semi- soft Amplitude Input layer Hidden layer Output layer 16 Figure 8. The topology of a MLP neural network used in almond nuts separation system in this study. Experiment 2 The whole stages of previous experiment were performed with six almond genotypes in three classes (Hard, Semi hard/ Semi soft and shell), each class included two genotypes (shown in Fig. 1). Based on the foregoing mentions, in this experiment conditions, 12 PE's in hidden layer had the least standard deviation error. By examining 17 different configurations, the best PC combinations for the input vector were found as 12 signal PSD (.1% of variance omitted), 27 amplitudes (.5% of variance omitted) and 7 phase angle features (.1% of variance omitted) so that the topology ( ) of MLP revealed the highest correlation coefficient (.89,.87 and.97) and least MSE (.458,.566 and.173) with correct classification rate of 92.11%, 9.13% and 97.2% for hard, semi soft and soft shell almond nuts respectively. Results showed that, using the designed acoustic system in combination with artificial neural networks has sufficient potentiality to separate almond nut varieties and is more accurate than fuzzy inference method which presented by Ebrahimi and Mollazadeh (21). Comparison of experiments 1 and 2 results indicated that by increasing almond varieties number for classification, the system accuracy is going to be reduced. CONCLUSION In this study a separation system, based on combination of acoustic detection and artificial neural networks, was designed and tested for classifying some of Iranian's almond nut genotypes based on their shell thickness and stiffness. This method has high accuracy and can be used for separation of almond genotypes. An online rejection system is needed to study the real separation accuracy. Moreover, the feasibility of mentioned method for grading of different almond varieties based on their kernel percentage is being conducted. REFERENCES Altuntas E, Gercekcioglu R, Kaya C. 21. Selected mechanical and geometric properties of different almond cultivars. International Journal of Food Properties. 13(2), Anonymous United States standards for grades of shelled almonds. United States Department of Agriculture _ Agricultural Marketing Service _ Fruit and Vegetable Division _ Fresh Products Branch, Washington D.C. Cetin AE, Pearson TC, Tewfik AH. 24. Classification of closed and open shell pistachio nuts using voice recognition technology. Transactions of the ASAE, 47, Ebrahimi E, Mollazadeh K. 21. Integrating fuzzy data mining and impulse acoustic techniques for almond nuts sorting. Australian Journal of Crop Science (AJCS). 4(5), FAO FAOSTAT database. (accessed: 8 Feb., 213). Haykin S Neural Networks: A Comprehensive Foundation. Prentice Hall, New Jersey. Kalkan H, Yardimci Y. 26. Classification of hazelnuts by impact acoustics. In Proceedings 16th IEEE signal processing soci ety workshop on MLSP. pp Kalkan H, Ince NF, Tewfik AH, Yardimci Y, Pearson T. 28. Classification of hazelnut kernels by using impact acoustic time-frequency patterns. EURASIP Journal on Advances in Signal Processing. 28 (1-11). 116

10 Intl J Agri Crop Sci. Vol., 6 (14), , 213 Karimi H, Navid H, Mahmoudi A Detection of damaged seeds in laboratory evaluation of precision planter using impact acoustics and artificial neural networks. Artificial Intelligence Research. 1(2), Madam B, Rahemi M, Mousavi A, Martinez-Gomez P Evaluation of the behavior of native Iranian almond species as rootstocks. International Journal of Nuts and Related Sciences. 2(3), Mahmoudi A, Omid M, Aghagolzadeh A, Borgayee AM. 26. Grading of Iranian's export pistachio nuts based on artificial neural networks. International Journal of Agriculture & Biology. 8(3), MathWorks MATLAB User s Guide. The MathWork, Inc. Omid M, Mahmoudi A, Omid MH. 21. Development of pistachio sorting system using principal component analysis (PCA) assisted artificial neural network (ANN) of impact acoustics. Expert Systems with Applications.37(1), Onaran I, Pearson TC, Yardimci Y, Cetin AE. 26. Detection of underdeveloped hazelnuts from fully developed nuts by impact acoustics. Transactions of the ASAE, 49(6), Pearson TC, Cetin AE, Tewfik AH, Haff RP. 27. Feasibility of impact acoustic emissions for detection of damaged wheat kernels. Digital Signal Processing, 17, Pearson TC. 21. Detection of pistachio nuts with closed shells using impact acoustics. Applied Engineering in Agriculture, 17(2), Rumelhurt DE, Hinton GE, Williams RJ Learning internal representations by back -propagation errors. Nature, 322, UNECE Standard DDP Concerning the marketing and commercial quality control of inshell almonds. United Nations. New York, Geneva