DESIGN AND IMPLEMENTATION OF DIGITAL IMAGE PROCESSING ON WEED CLEANING SYSTEMS IN PAVING BLOCK

• Main Author: Andy Rusman, Tobias
• Format: Final Project
• Language: Indonesia
• Online Access: https://digilib.itb.ac.id/gdl/view/51031
• Institution: Institut Teknologi Bandung

Currently, the care being taken to clean the gaps between paving blocks from weeds is using manual and chemical methods. Manual cleaning uses a kape / scraper tool to scrape off weeds that grow in the paving block gaps, while chemical cleaning uses sprayed herbicides. The use of kape for cleaning paving blocks requires strong physical strength and endurance because the paving block area is quite large, so there is a risk of physical injury and exhaustion. The use of herbicides can be done quickly and cover a large area, but could endanger human health and the environment through spray residue and contamination of water infiltration in the area around herbicide use. To overcome this problem, new kind of tools which do not endanger human health and the environment are needed to help clean up weeds in the paving block gaps The cleaning system developed is an automatic cleaner using laser heating. Laser heating system has high precision due to a small spot area, so it requires assistance of other systems to detect weeds and determine laser targets. The auxiliary system can be designed and implemented using a camera with digital image processing for detection and localization. Image processing system created using a combination of image reverse projection and object detection to estimate the real coordinates of weeds detected in the image. In addition to image processing, the processing unit sub-system is designed and implemented to regulate the software execution on the system. The behavior of the weed cleaning system on paving blocks is modeled by FSM (Finite State Machine) which has 3 main states, namely idle, scanning, and heating according to the system main function. The image processing system can perform reverse projection with a maximum deviation of 4.2 mm and detect weeds on paving blocks with mAP of 20.3% and an average detection time of 1.567 seconds on 58 evaluation images. The detection performance of the model is still below the mAP benchmark of 26.6%, this is because the paving dataset has very few training images (215 samples) so that the model lacks image samples to produce better detection results. In addition, the test results of the image scanning sub-system can functionally detect and estimate the presence of weeds at the real coordinates and the midpoint of the weed parts, but further testing using more samples is needed to determine the level of precision of estimation of the real coordinates of the bounding box center point regarding weeds, the accuracy of laser heating of weeds, and detection performance of weeds under actual conditions. In addition to the image processing sub-system, the processing unit sub-system can manage the system software execution according to the FSM-regulated behavior. Further testing is required to ensure the integration between the processing unit sub-system with the heating sub-system and heater holders as well as the FSM-like behavior