← Knowledge Hub

Computer Vision

From pixels to measurements: segmentation, features and detection.

Overview

Computer vision turns images into numbers: which pixels are plant, how big is a fruit, where is a weed. In agriculture most tasks are one of three kinds: segmentation (mark the pixels of interest), classification (name what is in an image) and detection (find and locate objects).

Key concepts

Colour spaces

RGB mixes colour and brightness. HSV separates hue from brightness, which helps for fruit ripeness, but it is not immune to light: in our study the detected size of ripe tomatoes shrank below about 3,000 lx.

Vegetation indices

Simple band arithmetic such as Excess Green (ExG = 2g − r − b on chromatic coordinates) separates green plants from soil with almost no training data.

Feature-based vs deep learning

Hand-made features (shape, colour histogram, texture) plus a classical classifier need few images; CNNs need thousands but learn the features themselves.

Evaluation

Report precision, recall and F1 on images that were not used for training, and look at the confusion matrix, not only one number.

Code examples

import cv2, numpy as np

img = cv2.imread("field.jpg").astype(np.float32)
b, g, r = cv2.split(img / img.sum(axis=2, keepdims=True).clip(1e-6))   # chromatic coordinates
exg = 2 * g - r - b                                                    # Excess Green index
exg8 = cv2.normalize(exg, None, 0, 255, cv2.NORM_MINMAX).astype(np.uint8)
_, mask = cv2.threshold(exg8, 0, 255, cv2.THRESH_BINARY + cv2.THRESH_OTSU)
cv2.imwrite("plant_mask.png", mask)

import cv2

hsv = cv2.cvtColor(cv2.imread("tomato.jpg"), cv2.COLOR_BGR2HSV)
# red wraps around hue 0, so two ranges; thresholds are examples, tune them on your own images
mask = cv2.inRange(hsv, (0, 100, 60), (10, 255, 255)) | cv2.inRange(hsv, (170, 100, 60), (180, 255, 255))
print("ripe pixels:", int((mask > 0).sum()))

Case study

Case studies from my work: how few images a feature-based classifier needs, and how illumination changes colour-based tomato detection.

References from my work