The accidental mixing of different grain types at an elevator is a common and costly mistake resulting in either a total loss, or costly blending processes to correct. We present an automated and intelligent quality assurance system to prevent these incidents by using a deep convolutional neural for crop type classification deployed to an edge device.
Our research shows human crop type classification accuracy across eight different crop types is 99.7% for the best expert and as low as 83.9% for the worst novice with an overall accuracy of 96.2% for the entire group in lab experiments. It is anticipated that in real world conditions when faced with fatigue, distraction, varied skill level and negligence this number would reduce further. It is clear that a highly accurate classification system which is able to continuously monitor the process and warn the operator would help reduce incidents.
Deep convolutional neural networks have shown excellent performance at many complex image classification tasks and a number of popular architectures have emerged (Inception V3, Google and ResNet Microsoft). However these architectures are very computationally and memory expensive making them challenging to deploy to low cost rugged hardware needed for many industrial application spaces. In addition the typical process of retraining only the final layers of the network (a technique called Transfer Learning) does not allow for customizing the network architecture to reduce complexity and potentially achieving better performance.
To overcome these limitations we have constructed a custom deep convolutional neural network architecture consisting of five convolution layers and two fully connected layers. The trained model achieves an overall accuracy of 99.5% (surpassing the human classifiers) and a Kappa coefficient of 99.48% across eight different crop types with one ? five cultivars (varieties) per crop type. With an added voting algorithm based on N samples the number of misclassifications can be further reduced and add robustness. The resultant optimized model at only 800KB (weights and biases) represents a reduction of two orders of magnitude with respect to storage and memory requirements enabling it to be run in real-time at the edge requiring no network connectivity or communication with the cloud.
