Agriculture and food

Computer vision for crops and produce

Use RGB and spectral imagery to investigate crop conditions, locate produce, and develop task-specific inspection workflows. Start with representative images and a measurable decision.

Choose the task

Match the workflow to the crop and decision

Locate and count produce

Detect or segment fruit in field images and video, then validate counts against ground truth.

Explore field imagery →

01 · Locate produce

Detect and count visible crops

Object detection and segmentation can locate fruit or other visible targets in images and video. Counts can help assess a workflow, but field estimates must account for occlusion, lighting, camera viewpoint, and sampling.

This apple detection clip is a demonstration, not a validated yield estimate. See the tomato and apple inference demos for Hailo HEF and ONNX comparisons; a production estimate requires validation against ground truth in the target crop and conditions.

Apple detection in an orchard using a Hailo HEF model

02 · Analyze crop quality

Combine visible shape with spectral evidence

Multispectral and hyperspectral images can reveal information beyond visible color. In Vision Studio, inspect bands and signatures, annotate regions, and train spectral classifiers or regressors where labeled data supports the task. Indices such as NDVI require suitable wavelength bands and metadata, and their value should be tested with representative measurements. Explore the spectral workflow.

A pipeline can locate items in an RGB view, then classify their corresponding spectral crops. This pattern is useful to investigate for sorting or quality tasks when surface appearance alone is insufficient. For rare defects, evaluate anomaly detection from normal images and use held-out produce examples; defect-region labels enable localization metrics. See the industrial inspection workflow.

A separate Studio example processes strawberries with instance segmentation and an ONNX spectral classifier. It demonstrates a workflow, not grading accuracy on new produce or production throughput; both require project-specific validation.

Vision Studio strawberry pipeline with instance segmentation and ONNX spectral classification
Instance segmentation followed by ONNX spectral classification; one sample frame produced seven detections. This is a workflow example, not a benchmark.

03 · Project evaluation

Prepare representative crop data

Useful starting material includes sample images or scenes, sensor details and wavelengths, lighting conditions, labels or reference measurements, and the decision the system must make. These help determine whether RGB, spectral imaging, or both are appropriate.

Bring your vision problem to us

Tell us about the crop, images or sensor, operating conditions, and decision you need to support. We can identify a practical first evaluation.