Article Preview
TopIntroduction
Weed control in fields involves identifying and characterizing the type of weed. As such, the automatic programmed method allows the recognition and classification of known herb types have important applications in this field. Based on the importance of classifying weeds in controlling weeds, researchers have proposed numerous algorithms to identify cultivated weeds (Pulido et al., 2017) (Forero et al., 2018). Researchers classified weeds using SVM with a recognition rate of 98% (Dos Santos et al., 2017). Others have used images to replicate the same experiments with other lenses. They implemented the automatic threshold and adaptive contour to segment parts, and used the method known as the smallest error for classifying the classes, and the recognition rate was on average 96% (Olsen et al., 2015). Other experiments used other algorithms based on K-nearest neighbors. However, this algorithm does not handle samples well, which are sometimes or always unbalanced. If the sampling capacity of one class is higher, while the sampling capacity of other classes is lower, it causes problems. To understand this concept well, when a sample that is new enters the system to be diagnosed, it can generate a class with a great capacity to be dominant in the K nearest neighbors of this new sample. Another influencing circumstance is that this algorithm is computationally expensive due to the categorization of each sample in order to calculate the distance of the sample from all other known samples in order to obtain its K nearest neighbors. Nevertheless, from another point of view, this algorithm is more adequate for larger data samples, and for data with few samples, this approach is more likely to may produce misclassified classes. Whether the result is interesting or not, these experiments only use a limited quantity of images, or use their database with modifications made to test their identification, after these manipulations the results become biased (Muppala & Guruviah, 2020). To our knowledge, there are no large public databases for the detection and classification of weeds that have spread in the environment of our study. Therefore, in order to identify the real effectiveness of the proposed algorithm objectively, we collected a set of real field images taken by our professional camera and combined them with images that we obtained on from public dataset released with free licenses. This dataset is filtered and preprocessed, to generate over 3000 images (comprising a large portion of 75% training data and 25% test data). In these datasets, we divide the weeds into 4 classes according to the types of these weeds, which are; Beta vulgaris subsp, Capsella, Chenopodium, Galium aparine (Jabir et al., 2021). These weeds are the most common in the study environment.
With the development of characteristics of computers, which improve its performance, the convolutional neural network has become more powerful and popular in different problems; various improvements of the basic CNN architecture have been made from the nineties to the present day. These changes for the purpose of improvement can be distinguished between regularization, optimization of parameters, structural reformulation, etc. However, it can be said that the goal of improving the performance of CNN comes from reformulating and processing units and building new blocks. Most of the improvements in architecture of CNN have been made on depth and space. The types of innovations made architecture can be classified into seven different categories, namely; spatial exploitation, depth, multipath, width, characteristic map exploitation, channel amplification, etc (Milioto et al., 2018), among the CNN models which interest us a lot in this weed detection experiment, we cite, ResNet (Mahajan & Chaudhary 2019), VGG and ResNeXt (Hitawala, 2018).