You can deploy this model and start earning money today!

ai-department-lpnu/
YOLOv8-Ghost-QWA

Name: YOLOv8-Ghost-QWA
Availability: InStock

agriculture

tomato

object-detection

instance-segmentation

Ultralytics/YOLOv8

finetune

cc-by-4.0

Views

No views yet

Tomato Ripeness Segmentation with YOLOv8-Ghost-QWA

To use YOLOv8-Ghost-QWA, you need to follow the instructions below:

Install the modified Ultralytics library. For that, run the below command:

pip install git+https://github.com/nazarkohut/tomato_ripeness_enhancement.git

Next, download the model weights best_yolov8_ghost_qwa.pt from this repository. You can also download an example image for inference, or use your own.

Example image:

Then use the following code snippet, passing the path to your local image file:

python

from ultralytics import YOLO
import matplotlib.pyplot as plt

image_path = "/path/to/local/image.png"
yolov8_ghost_qwa = YOLO("models/best_yolov8_ghost_qwa.pt")
results = yolov8_ghost_qwa.predict(image_path, conf=0.4, verbose=False)
annotated_bgr = results[0].plot(boxes=True)
annotated_rgb = annotated_bgr[:, :, ::-1]    # flip channels

plt.figure(figsize=(10,6))
plt.axis("off")
plt.imshow(annotated_rgb)
plt.show()

Alternatively, you can use OpenCV to read the image and pass it as a numpy array:

python

import cv2
from ultralytics import YOLO

model = YOLO("models/best_yolov8_ghost_qwa.pt")

image_path = "/path/to/local/image.png"
img_bgr = cv2.imread(image_path)
results = model.predict(source=img_bgr, conf=0.25, verbose=False)
annotated_bgr = results[0].plot(boxes=True)
annotated_rgb = annotated_bgr[:, :, ::-1]

plt.imshow(annotated_rgb)
plt.axis("off")
plt.show()
print("boxes:", len(results[0].boxes))

Model architecture

The main improvement over the original YOLOv8 is the use of Ghost Convolution modules to accelerate inference, especially in constrained or embedded environments. Since lightweight convolutions can reduce representational capacity, the model employs SimAM attention in the backbone and QWA attention in the head (after each C2F module) to retain strong feature quality. These attention components effectively offset the potential accuracy drop, ensuring performance remains close to the standard YOLOv8 while delivering higher efficiency.

Dataset

The dataset used to train this model consists of images from LaboroTomato and our ITomatoSegUa: merged-laboro-and-itomato-seg-ua

ai-department-lpnu/YOLOv8-Ghost-QWA

Tomato Ripeness Segmentation with YOLOv8-Ghost-QWA

Model architecture

Dataset

ai-department-lpnu/
YOLOv8-Ghost-QWA