Lightweight L-YOLO Model Enhances High-Resolution Small Object Detection Efficiency

Researchers Kaiyun Luo and Runmiao Zhou have introduced L-YOLO, a lightweight computer vision model designed to enhance small object detection in high-resolution imagery while reducing computational overhead. Published in Discover Applied Sciences, the study demonstrates that integrating a MobileNetV2 backbone with a dedicated multi-level fusion attention module allows the system to preserve fine-grained structural details during inference without imposing prohibitive resource demands.

High-precision detection of small items traditionally presents an engineering dilemma in computer vision. As input resolution increases, the required computation and memory occupancy expand substantially. Conversely, conventional approaches that simplify network complexity to conserve processing power frequently sacrifice delicate, low-level feature representations that are essential for identifying tiny targets.

To address this balance, L-YOLO incorporates a backbone network based on MobileNetV2 coupled with a small object multi-level fusion attention module. The architecture retains subtle details under high-resolution processing. To ensure experimental reproducibility, the authors aligned the training objective with the native YOLOv8 formulation and evaluated the models under a standardized benchmarking protocol on the COCO dataset using a batch size of one.

Under the unified 448 by 448 resolution evaluation protocol on COCO small objects, L-YOLO demonstrated clear quantitative gains over the baseline. The model increased the small-object average precision metric, APsmall, from 14.3 percent to 16.5 percent, and improved AP50-small from 27.1 percent to 31.2 percent. These accuracy gains were achieved alongside computational savings, reducing floating-point operations from 6.0G to 4.2G and lowering end-to-end inference latency from 11.1 milliseconds per image to 9.8 milliseconds per image.

The investigators also conducted hierarchical experiments across multiple image resolutions, testing inputs at 640 by 640, 1024 by 1024, and 1280 by 1280 pixels. These tests indicated that detection efficiency gains were particularly pronounced for smaller targets at higher resolutions. Notably, at an operating resolution of 1024 by 1024 pixels, the system achieved an actual processing throughput of 41.2 frames per second while securing a 39.3 percent APsmall improvement over baseline.

Ablation analyses conducted by removing individual modules confirmed that incorporating attention mechanisms contributed gains that exceeded simple weight reduction, successfully transforming subtle high-resolution features into effective low-level object representations at a sustainable computational cost. Beyond benchmark evaluations, the researchers validated the practical utility of the framework through a case study monitoring a fishing vessel, which confirmed measurable gains in positioning accuracy.

Subject of Research: Lightweight neural network architecture for high-resolution small object detection

Article Title: Multiscale attention enhanced lightweight YOLO for high resolution small object detection

Article References: Luo, K., & Zhou, R. (2026). Multiscale attention enhanced lightweight YOLO for high resolution small object detection. Discover Applied Sciences. https://doi.org/10.1007/s42452-026-09828-y

Image Credits: AI Generated

DOI: 10.1007/s42452-026-09828-y

Keywords: L-YOLO, YOLOv8, MobileNetV2, object detection, small object detection, high-resolution reasoning, multiscale attention, multi-level fusion attention, COCO benchmark, inference latency, FLOPs reduction, computational efficiency, throughput, ablation study, fishing vessel monitoring, position accuracy, feature retention