Abstract:
Agriculture is the origin of human sustenance in this world. The detection and classification of weeds are of the most important technical and economical importance in the agricultural industry. A weed is a plant growing along with useful agricultural products. A weed may be defined as any plant or vegetation that interferes with the objectives of farming or forestry, such as growing crops, grazing animals, or cultivating forest plantations. This weed decreases the growth of the crop and reduces the farm yield; hence these weeds should be identified and classified. The classification of weeds is helpful for adopting weed management methods for a particular group of weeds instead of against an individual weed species.
In past days weed detection was done by employing some men, particularly for that intention. In the olden days, weed detection was done by inspecting each and every place in the field. Then weeds were manually removed. Later with the improvement in technology, people started using herbicides to take out the weeds. But to identify the weed still physical power was used in many parts of the world. Later there came few methods to discover the weeds without human intervention but due to lack of their accuracy, they were incapable to reach the public. Then image processing was used for this purpose. In this project, I developed a system based on deep learning and image processing that helps to classify different types of weeds, which have become a major threat to our vegetable plantation. The idea for this project came from an IEEE paper I read. That paper suggests Convolutional Neural Networks for developing such a system. I also referred to another IEEE paper that uses centernet architecture and genetic algorithm to deal with this. Of the two, I choose another method ie, vgg16 of CNN to develop the project.
So here I am using the vgg16 architecture of CNN to iidentifyingand classify weeds. There are 09 classes of weeds. To train the system, I used the dataset from Kaggle. The DeepWeeds dataset consists of 17,509 unique 256x256 color images in ‘.png’ format, around 1000 for each of the 9 classes. There are 15,007 training images and 2,501 test images. These images were collected in situ from eight rangeland environments across northern Australia.
