Improving Weeds Detection in Pastures Using Illumination Invariance Techniques

Computers have various applications in relation to the classification of weeds, including computer vision. This paper demonstrates the use of illumination invariance techniques and shadow reduction in images to improve the accuracy of machine learning (ML) models using support vector machines. The paper’s main aim is to identify the benefits of image optimisation utilising adjusting dark images. More specifically, the paper uses brightness and contrast adjustment to fix images and then compares the results of a dataset that underwent image pre-processing and a dataset that did not. Ensuring the clearness of an object in an image is essential if a ML model is to identify it accurately. Many issues within image datasets can hinder the accuracy of ML classification models, for example, illumination invariance and shadowed images, which entail underexposed dark pictures being projected onto the target in the absence of light sources. The paper uses several techniques and technologies to analyse the data, including cross-validation, a confusion matrix, the pre-processing technique, the TensorFlow framework and training conducted in both the central processing unit and graphics processing units. The results of these analyses show that the brightness has significantly enhanced the accuracy of the ML model. In addition, applying image pre-processing to the shadow has resulted in a slight improvement of 1% in this regard. In conclusion, this paper presents evidence concerning ML-based solutions for improving the accuracy of classification models by enhancing images of weeds using pixel brightness transformations.