Multiscale attention enhanced lightweight YOLO for high resolution small object detection

High-precision small object detection is usually faced with the problem that the higher the input resolution, the greater the amount of computation and memory occupation, but reducing the complexity of the model will lead to the loss of detailed features. L-YOLO designed in this paper is a lightweight YOLO model, which is integrated into the backbone network based on MobileNetV2 and the small object multi-level fusion attention module, so as to better retain the details of small objects under high-resolution reasoning. The training objective follows the native YOLOv8 formulation to maintain reproducibility, and the models are evaluated under a unified protocol on COCO with batch size 1. On COCO small objects, L-YOLO improves APsmall from 14.3% to 16.5% and AP50-small from 27.1% to 31.2%, while reducing FLOPs from 6.0G to 4.2G and end-to-end inference latency from 11.1 ms/image to 9.8 ms/image under the unified 448 × 448 protocol. In the hierarchical experiment with 640 × 640/1024 × 1024/1280 × 1280 resolution, it can be found that the detection efficiency is greater for small objects. At 1024 × 1024, we achieve higher actual throughput (41.2 FPS) and larger APsmall(39.3%) improvement over baseline. After removing some modules to prove its contribution, we find that the addition of attention improves APsmall and exceeds the simple weight reduction effect, so that the subtle features in high-resolution images can be converted into effective low-level object information while maintaining an appropriate cost consumption. In addition, a case study of monitoring a fishing vessel was carried out in detail, and the improvement of position accuracy was provided.

Authors

Institutions

Publication Details

Journal
Discover Applied Sciences
Published
2026-11-13
DOI
https://doi.org/10.1007/s42452-026-09828-y
Primary Topic
Advanced Neural Network Applications
Type
article
Field-Weighted Citation Impact
0.00
Controls
|||
ALL TIME
JAN
FEB
MAR
APR
MAY
JUN
JUL
AUG
SEP
OCT
NOV
article

Multiscale attention enhanced lightweight YOLO for high resolution small object detection

Kaiyun Luo, Runmiao Zhou
2 links
Discover Applied Sciences
Advanced Neural Network Applications
article

Multiscale attention enhanced lightweight YOLO for high resolution small object detection

Kaiyun Luo, Runmiao Zhou
article en

Abstract

High-precision small object detection is usually faced with the problem that the higher the input resolution, the greater the amount of computation and memory occupation, but reducing the complexity of the model will lead to the loss of detailed features. L-YOLO designed in this paper is a lightweight YOLO model, which is integrated into the backbone network based on MobileNetV2 and the small object multi-level fusion attention module, so as to better retain the details of small objects under high-resolution reasoning. The training objective follows the native YOLOv8 formulation to maintain reproducibility, and the models are evaluated under a unified protocol on COCO with batch size 1. On COCO small objects, L-YOLO improves APsmall from 14.3% to 16.5% and AP50-small from 27.1% to 31.2%, while reducing FLOPs from 6.0G to 4.2G and end-to-end inference latency from 11.1 ms/image to 9.8 ms/image under the unified 448 × 448 protocol. In the hierarchical experiment with 640 × 640/1024 × 1024/1280 × 1280 resolution, it can be found that the detection efficiency is greater for small objects. At 1024 × 1024, we achieve higher actual throughput (41.2 FPS) and larger APsmall(39.3%) improvement over baseline. After removing some modules to prove its contribution, we find that the addition of attention improves APsmall and exceeds the simple weight reduction effect, so that the subtle features in high-resolution images can be converted into effective low-level object information while maintaining an appropriate cost consumption. In addition, a case study of monitoring a fishing vessel was carried out in detail, and the improvement of position accuracy was provided.

Discover Applied Sciences
Hunan University of Arts and Science (CN), The First People's Hospital of Changde (CN)
Advanced Neural Network Applications
AI Navigator

Ask Laika to Summarize, Analyze, and Connect papers live on the map.

Summarize Papers & Methodologies

Extract key findings, datasets, and comparative methods across publications.

Benchmark Rankings & Visual Analytics

Rank top research institutions, authors, funders, topics, and journals by Field-Weighted Citation Impact (FWCI) and paper volume with instant charts.

Connect Distant Disciplines

Bridge topological clusters on the map to find hidden collaborative intersections.

Laika access coming soon