Deep W2net: Deep learning based YOLO network for accurate weed detection in wheat fields

Aim of study: To detect weeds in the wheat-growing regions of northwestern India using a novel Deep W2net to enhance wheat protection. Area of study: Northwestern India. Material and methods: The captured video sequence was transformed into frames, and the frames were denoised to enhance image clarity by sharpening edges and reducing blurriness. The denoised images were augmented to increase the size of the dataset for training. The proposed Deep W2net consists of Squeeze and Excitation (SE) and Swin Transformer (ST) modules integrated YOLO network (SET-YOLO) for efficient detection of weeds. The SE adjusts the dependencies between convolutional kernel channels, and ST enables the model to analyse spatial relationships at multiple scales. Main results: The proposed Deep W2net attains an accuracy of 99.16% for detecting the weeds in wheat fields. The proposed model improves the range of overall accuracy by 4.19%, 1.47%, and 6.51% better than the existing techniques such as CNN, CSCW YOLOV7, and YOLOV8 respectively. Research highlights: The proposed Deep W2net method demonstrates a significant advancement in weed detection to maximize the wheat productivity and potentially lead to more efficient and precise crop management.