Seed Quality Assessment 

Hello guys today we  will emphasize on the importance of seed quality in ones life and the way to implement the seed quality assessment. In the commercial seed cultivation sector, it is not feasible to filter out the individual damaged seeds and foreign elements in the case of the voluminous amount of seeds. Moreover, in the seed packaging industries, the seed filtering is done by the manual labor force which leads to poor quality seeds in the packets and waste of a large amount of fine quality seeds. In this context, efficient and automated seed testing is the most important part of all other seeds technologies.

We will be implementing Convolutional Neural Network (CNN) for detecting the quality of the image Seed Lot (16 seeds) as Excellent, Good, Average, Bad and Worst quality. CNN is an extension of the Artificial Neural Network(ANN) that is comprised of neurons that optimize itself through continuous learning [13]. In the paper [13],CNN composed of 3 layers viz. convolution layers, pooling layers and fully connected layers. They implemented it for the MNIST dataset for detecting the hand-written digit

The system can be developed in four stages: Dataset Preparation, Image Pre - processing,

Building CNN, and Compiling and Training CNN.

To generate our dataset, we used the maize variety, Super 900 M - F1, which is largely available.

We included fine seeds, damaged seeds and foreign elements in the dataset and categorized the seed lot in percentage basis (Table 1) as Excellent, Good, Average, Bad and Worst as in figure 1. First, we clicked the images of the seed lot with a normal camera with not so good resolution. But the detailing that the system require to distinguish between the fine seed and damaged seeds, were not good enough which could make our system vulnerable to inaccurate predictions. So we again clicked 3000

images with high-quality camera and generated high-quality dataset with sharp detailing. We tried to

accommodate all the possible orientations of the seeds in the seed lot so that the model learns all the

different positions of the seeds in the cluster. Finally, we separated the dataset as 2500 training data and

500 testing data belonging to five categories

The dataset should be processed before feeding it into the neural network. We augmented the dataset using the library function provided by Keras and processed all the images to uniform 256 *256 resolution. Building a neural network was a great challenge because we need to make sure that in any of the layers, the detailing of the damaged and fine seeds is not lost. So we built our CNN keeping the high resolution for the input images. After the model generation, we tested our system for the 20 test images in GUI developed using Django Framework.

Seed Quality assessment will minimize the manual labour and also ensure that there is no manual mistake thereby giving high quality assurance