Introduction: Farmers use the traditional method (unaided eye and dimensional examination of the seeds) to determine the percentage of healthy rice paddies after harvesting and drying. In recent years, however, machine vision has grown significantly and many researchers have used this system to classify and measure the quality of agricultural and food products, identify different varieties of products, determine the geographical origin of products, identify losses, estimate effective substances, and other related subjects. For example, in one study, the mixing percentage of wheat grains and the determination of grain hardness were performed with the help of machine vision (Fazaeli and Afkari-Sayah, 2009). In order to calculate the hardness of seeds, they used the weight percentage and to measure the mixing percentage, they used the color characteristics of the samples. The classification accuracy obtained in this research was 93%. In another research, artificial neural networks (ANN) and some statistical techniques were used to identify and classify 5 varieties of Iranian rice (Hatami et al., 2010). In their study, samples of 300 seeds were randomly selected from each cultivar, and images of the samples were captured using a digital camera. Then, using the thresholding process, the images were segmented and the features of the area, length of major and minor axis, perimeter, convexity level, strength, extent, and equivalent diameter were extracted. In the created regression model, the highest detection percentage was obtained in Garde and Ramadani cultivars equal to 93.6 and 91.48%, respectively. According to the previous studies, it was found that determining the percentage of crop loss and the percentage of paddy harvested from the field before the bleaching stage is very important so that the percentage of paddy harvested from the field shows the yield of the crop per hectare. Therefore, in the current research, an intelligent machine vision system was developed and evaluated to quickly detect and determine the percentage of healthy paddy seeds. Materials and methods: First, five varieties of rice paddy were prepared. The samples included devoid and healthy seeds. In order to determine the moisture content of the samples, the samples were placed in an oven with a temperature of 70ºC for 24 hours. Next, the images were captured online with the help of a 16-megapixel camera that was placed perpendicular to the surface of the samples in the chamber of the imaging system. The distance between the camera and the samples was fixed at 30 cm. In addition to the imaging camera, the system includes a cubic enclosure made of stainless steel and a lighting unit system. Two projector light sources were installed in the upper half of the chamber so that proper and completely uniform lighting (using optical filters) was guaranteed. To prepare pictures, 250 seeds were randomly selected from healthy samples and pictures were taken. Then the images containing the devoid seeds were taken according to the previous settings and under the same lighting conditions. Moreover, in order to evaluate the algorithm, samples were taken and stored as a combination of healthy and devoid rice with 5%, 10%, and 15% devoid percentages. Next, the machine vision system algorithm was designed and coded in MATLAB software. In the order that the images were read first in the software, pre-processing was applied to the images and finally, their color and shape characteristics were extracted. Finally, a multilayer perceptron (MLP) artificial neural network with the Levenberg-Marquardt algorithm was used for clustering and separating full and devoid seeds. The input layer of MLP had the number of neurons equal to the number of features considered in each stage of training. In the present study, a neural network with a hidden layer was used, the number of neurons of which was selected as 10 by trial and error. Results and discussion: First, the color characteristics of healthy and devoid rice were extracted. Due to the non-uniformity of the results obtained from the analysis of color channels, PCA analysis was used to check the discriminating power of color features. For example, in the Binam cultivar, R color channels related to healthy and devoid seeds overlapped and it was not suitable for separating and clustering the seeds. For this variety, color channel B was chosen as a suitable option for separating or distinguishing between healthy and devoid seeds. For the Neda variety, the R color channel was very clearly separated for healthy and infected samples. In the second stage, the PCA analysis method was used to determine the more effective features that play a more important role in identifying healthy seeds from devoid ones. The results of the PCA analysis showed that by using all the shape and color features, a high percentage of mixing between healthy and devoid samples was created and it is not possible to use all the features to separate healthy and devoid samples in this way. Due to the fact that it is not possible to recognize the shape characteristics of rice grains when they are moved by volume on the conveyor belt, PCA analysis was performed using 3 extracted color features, where it is possible to check the image data pixel-by-pixel and to evaluate the operational usability of the system in online conditions. PCA analysis graphs, considering only 3 color characteristics of healthy and devoid rice samples, show that in the PCA score diagrams, healthy and devoid data can be separated to an acceptable extent and are separated in a certain range. Also, in the PCA loading diagrams, it can be seen that in all the tested samples, the color features are located at the farthest position from the center. This shows the ability of high resolution by all color features. The results of the training, validation, and testing of the perceptron neural network model created with 8 input neurons, one hidden layer with 10 hidden neurons, and two output linear layers were used to distinguish between devoid and healthy samples for different varieties of rice. According to the obtained results, the R² coefficient in training and final evaluation for the Roshan variety was equal to 0.86. In the same way, the highest R² coefficient in the final evaluation of the models created based on color characteristics was obtained for the Neda cultivar equal to 0.96. Conclusion: In this research, an intelligent system based on image processing methods and artificial neural networks was implemented and evaluated to identify the devoid rice seeds from five newly improved rice cultivars. The system proposed in this research has high accuracy and speed compared to conventional experimental methods and scientific methods that have been used so far. PCA method was used for a more detailed examination of the effective components in distinguishing between healthy and devoid seeds. Due to the need for the system to be online and the quality measurement of rice grains in piles, three color features were selected from among the features and modeled using MLP artificial neural networks. The final results, based on the statistical parameters, showed that the implemented system can be used as an online, fast, and cheap system for measuring the quality of paddy in the stages of entering the lines of rice bleaching factories. [ABSTRACT FROM AUTHOR]