Convolutional Neural Network Optimization for Deep Weeds

Abstract

Precision agriculture is critical in ensuring food availability while maintaining environmental sustainability. Weeds are a serious threat to crops because they can inhibit plant growth and absorption of nutrients and infect nearby plants. Reduction in agricultural production can reach 20-80% if weeds are not handled quickly and precisely. In this study, four Convolutional neural network architectures were implemented to identify weeds based on images. The total number of images in the dataset used is 17,509 images grouped into nine classes which are divided into 80% for training data and 20% for test data. The training process uses a transfer learning scheme and operates several different optimization functions. The test results show that the best performance is achieved by the GoogleNet architecture using the stochastic gradient descent with momentum optimization function with a classification accuracy of 92.38%. Testing also shows that the ShuffleNet architecture classifies images faster than the other architectures used in this study, although its performance is slightly lower than GoogleNet.

A Fine-Tuned RetinaNet for Real-Time Lettuce Detection

“…Object recognition is crucial, yet the two differ significantly [4]. The primary difference is that object detection focuses on determining the location of objects using bounding boxes [5]. In contrast, object recognition takes an additional step by categorizing and labeling the identified objects based on predefined classes [6].…”