Easy Ways to Boost YOLO: Systematic Review of Versions and Use Cases

Overview

The author surveyed recent YOLO literature, including the newly released YOLOv13 and upcoming YOLOv14, compiling 113 papers and more than 20 defect‑detection datasets. Improvements are organized into five categories based on technical essence, mechanism, and scenario suitability, enabling both scholars to structure their papers and developers to pinpoint optimization stages.

Architecture Enhancement Components

Key ideas: adding attention mechanisms, integrating Mamba, and multi‑scale feature fusion to enlarge receptive fields and improve context modeling.

Reference: SCCA‑YOLO: A Spatial and Channel Collaborative Attention Enhanced YOLO Network for Highway Autonomous Driving Perception System .

Method: Introduces a spatial‑channel collaborative attention module combined with Ghost modules for efficient computation.

New spatial‑channel collaborative attention improves feature expression.

Ghost module reduces computational cost.

Constructed a large animal dataset for rural road scenarios.

Model Efficiency Optimization

Key ideas: lightweight backbones, knowledge distillation, and attention to achieve real‑time inference on edge devices.

Reference: YOLO‑Granada: a lightweight attentioned Yolo for pomegranates fruit detection .

Method: Uses ShuffleNetv2 as backbone and CBAM attention, cutting model size and FLOPs while preserving accuracy.

ShuffleNetv2 backbone dramatically reduces size and computation.

CBAM attention enhances feature extraction.

Inference speed improves by 17.3% compared with the original network.

Multi‑Task Collaborative Learning

Key ideas: combine segmentation (SAM) with detection to share backbone features and boost semantic understanding.

Reference: Intraoperative Glioma Segmentation with YOLO + SAM for Improved Accuracy in Tumor Resection .

Method: YOLOv8 provides fast tumor localization; SAM refines segmentation, achieving a Dice score of 0.79 on noisy MRI and inference time of 15‑25 s.

YOLOv8 + SAM yields rapid detection and precise segmentation.

Robust training on noisy MRI images.

Inference time suitable for real‑time surgery.

Temporal Modeling and Filtering

Key ideas: integrate Kalman filter to smooth detection trajectories across video frames.

Reference: Seedling maize counting method in complex backgrounds based on YOLOV5 and Kalman filter tracking algorithm .

Method: Enhanced YOLOv5 (SE‑YOLOV5m) with channel attention, then applies Kalman filter for tracking and counting, adding baseline and threshold parameters to mitigate edge distortion and missed detections.

SE‑YOLOV5m improves accuracy in complex backgrounds.

Kalman filter prevents duplicate counts across frames.

Baseline and threshold handle edge cases.

Domain‑Specific Customization

Key ideas: tailor loss functions, data augmentation, and attention modules for industrial defect detection.

Reference: A high precision YOLO model for surface defect detection based on PyConv and CISBA .

Method: Combines EMA, PyConv, CISBA, and Soft‑NMS to enhance small‑object detection on noisy surfaces, achieving high precision and real‑time speed.

EMA multi‑scale attention improves focus on varied object sizes.

PyConv pyramidal convolutions boost multi‑scale feature extraction.

CISBA attention strengthens detection in complex backgrounds.

Resources

All referenced papers include links to open‑source code repositories, and the author provides a compiled collection of 113 YOLO papers, code bases, and datasets for further study.