Choose an application

• Reference Manager
• EndNote
• RefWorks (Direct export to RefWorks)

Phonetics

Choose an application

• Reference Manager
• EndNote
• RefWorks (Direct export to RefWorks)

Choose an application

• Reference Manager
• EndNote
• RefWorks (Direct export to RefWorks)

In de landbouw is er nood aan een systeem die een halt toe roept aan de verspilling van pesticides. Dit komt niet alleen ten voordele van de natuur maar ook van de portemonnee van de landbouwer zelf. Om dit te verwezenlijken is er een toepassing nodig die in staat is onkruid te herkennen en direct plaatselijk te verdelgen. Daarvoor is het zeer belangrijk dat er een grote datacollectie van onkruid in alle mogelijke omstandigheden (regen, volle zon, sneeuw, bewolkt, …) wordt gemaakt. Zodat uit deze datacollectie nauwkeurige algoritmes kunnen geschreven worden. De bachelorproef gaat over een systeem ontwerpen om deze datacollectie aan te maken. In dit project moeten hoge resolutie camera’s beelden maken van onkruid in alle omstandigheden. Deze beelden dan linken aan een datum, gps locatie en een gewassoort. Waarna de beelden door een andere firma kunnen gebruikt worden om algoritmes te schrijven voor onkruidherkenning.

Choose an application

• Reference Manager
• EndNote
• RefWorks (Direct export to RefWorks)

Choose an application

• Reference Manager
• EndNote
• RefWorks (Direct export to RefWorks)

Potatoes (Solanum tuberosum) are one of the most cultivated crops in Belgium. Albeit potato plants grow in a wide range of environmental parameters, they are highly susceptible to diseases. One of the devastating diseases is called early blight and is caused by the fungi Alternaria solani. Hence, potato producers are compelled to use fungicides between 10-25 times per year to avoid losses. Excessive use of fungicides is correlated with a selection process of pesticide resistant species. Pesticide applications can be reduced by a targeted application when and where necessary, based on principles of precision agriculture. This however requires effective lesion detection techniques. The deployment of novel techniques based on convolutional neural networks (CNN) in combination with high resolution RGB imagery seems to be a promising solution. Therefore, the deployment of Mask R-CNN, a state-of-the-art instance segmentation model, has been studied using transfer learning on RGB images of potato plants acquired in a greenhouse. This thesis had three objectives: Firstly, determining the optimal setup of hyperparameters such as optimizer, learning rate, layers to be (re-)trained and number of epochs until convergence. The findings suggest, using ResNet-50 as a backbone with pre-trained MS COCO weights, to train only the heads, using the Adam optimizer, initial learning rate of 0.002, for 88 epochs and a minimum detection confidence of 0.9. Secondly, assessing the robustness of the results by exposing the model on distorted images common to the domain of agriculture (e.g. change of coloration due to changing illumination). Certain distortions, such as decreased brightness, change of color temperature, Gaussian blur and color contrast enhancement, yielded a higher precision, but a lower recall, indicating an overall decrease in numbers of instances detected. The remainder of the distortions, such as JPEG compression quality, resizing, added noise and increasing brightness showed a more or less synchronized trajectory of scores. The sensitivity of the scores with respect to increasing degrees of distortion was assessed by the means of the F1-score. If the F1-score undercut 7.5 % of the score on the original test set or if the distortion type was expected to be beneficial for the training, it was selected to be used as a candidate data augmentation technique. Thus, the third objective used the insights from the previous objectives to create a D-optimal experimental design screening of significantly positive augmentation techniques. Therewith the model has been trained with four scenarios: maximizing precision, maximizing recall, using all techniques and none. The scenario maximizing precision led to an overall decrease of detected instances, yet scored the highest precision with 0.81, 0.05 better than without data augmentation. The scenario maximizing recall has failed its aim and the recall score was maximized by using all techniques, yielding a recall of 0.65, also 0.05 better than without data augmentation. A visual inspection of the model inferences further showed that the model had difficulties detecting smaller lesions or lesions with added features, as for instance scar tissue, suggesting to scrutinize the labelling protocol and performance for detecting small objects using Mask R-CNN.