Scene Recognition TRIZ Cases

A same-image trained classifier improves concrete and asphalt detection in remote imagery despite atmospheric, regional, and calibration variation.

Switching ANN resolution by scene visibility cuts edge-camera compute load in fog or smog while preserving image analysis accuracy.

Image-based passenger detection lets elevators skip empty calls, manage car capacity, and support social distancing with fewer delays.

Behavior prediction is made more interpretable by combining metadata encoding, self-focused attention maps, and visualization of factors affecting results.

Hybrid DDM-AHP weighting improves geospatial hazard vulnerability mapping, reducing field surveys while capturing slope instability risk.

Lidar obstacle points periodically tune radar filters to cut false detections and improve obstacle sensing in adverse weather.

Multiple image-difference metrics, PCA, and clustering pre-select satellite image pairs with likely changes, cutting manual review time.

Multiple image-difference metrics are reduced and clustered to flag satellite image pairs with significant changes before manual review.

Self-supervised embedding feedback trains object detectors with far less manual labeling, cutting bias, cost, and training time.

Height-based cloud segmentation separates stacked layers so each advection vector can be estimated more accurately from visible and infrared images.

Neural networks combine satellite, weather, and agronomic data to predict sugarcane biomass and sugar yield for better harvest timing.

Graph-based matching uses alternating trees and matrix multiplications to cut memory overhead and speed object detection on parallel processors.

Temperature-classified infrared training images let vehicles switch among specialized AI models to detect pedestrians despite low thermal contrast.

Uses model-generated boxes and confidence data to retrain image detection models for new domains with less manual annotation.

Visible and thermal object detection guides gain or offset changes when saturated pixels rise, preserving far-infrared temperature resolution.

Appearance-based clustering and anomaly scoring speed object labelling in geospatial imagery while helping analysts catch unusual targets.

Broadband FM radar with SAR imaging enables real-time runway scanning to detect small debris and pavement cracks with uniform high resolution.

Multi-clustering isolates sparse pedestrian and bicycle point clouds from interference for accurate real-time LIDAR localization.

Pairs point-cloud reflectance data from different wavelengths to predict missing spectral intensity and improve localization and mapping accuracy.

Avoids sunlit glare zones in the traffic light image so vehicles can measure signal color more reliably under harsh lighting.

A learned spectral weight matrix boosts foreground saliency and suppresses background clutter to cut false alarms and compute on low-SWaP platforms.

A reference-image correlation structure corrects neural network image predictions for adversarial and edge cases without costly retraining.

Dense depth maps and stability scoring isolate static scene points, enabling faster and more accurate image annotation for ML training.

Dedicated optical flow hardware and a vision processor track many feature points with subpixel precision while lowering real-time compute load.

Corner-point 3D bounding boxes improve occluded object detection and tracking when point clouds are incomplete in autonomous driving.

Adjacent vehicle position data is fused with ego localization to improve autonomous vehicle positioning without point cloud exchange or rigid body simulation.

Attention-based neural alignment links current and historical entity features to preserve object identity across time without ground truth labels.

Style-transferred source images and generated target annotations cut labeling effort while preserving object detection accuracy across domains.

Radar-track similarity and classification checks remove redundant video tracks, cutting false alarms from related objects in surveillance.

Camera-based anomaly detection classifies road events and routes location-tagged alerts to the right authority without distracting the driver.

Fiducial marker constellations help UAVs distinguish the correct charging pad despite limited GPS accuracy and marker placement variation.