Improved Feature Extractor and Scaling on the YOLOv7-Tiny Model to Improve Weed Detection Performance in Corn Plants

This paper was not presented at the conference.

Abstract

This paper was not presented at the conference.

Weeds in maize fields can significantly impact crop growth and yield, therefore making their accurate identification and management crucial for effective agricultural practices. This study aimed to enhance the performance of the YOLOv7-tiny model by developing a feature extraction technique specifically for detecting weeds in maize fields. Modifications to the backbone layer, including residual blocks and SPP modules, were implemented to improve the model's ability to detect and classify weeds amidst the maize plants. The evaluation results of the different models showed their effectiveness in detecting weeds in maize fields. The best YOLOv7-Tiny model showed the highest achieved FPS and competitive MAP scores, demonstrating its potential for accurately identifying and localizing weeds within the crop. The model achieved a MAP of 0.523 (IoU 0.5) and 0.22 (IoU 0.95) specifically for the weed object, indicating its proficiency in weed detection tasks. The findings of this study have implications for precision agriculture and weed management strategies in maize fields. By employing advanced computer vision techniques like the YOLOv7-tiny model, farmers and agronomists can better monitor and control weed populations, leading to improved crop health, increased yields, and reduced reliance on herbicides.