Technical Summary: Lightweight and Resolution-Flexible YOLO for Edge Devices (RLPF-YOLO)

1. Problem Statement

The deployment of lightweight object detectors on Unmanned Aerial Vehicles (UAVs) faces a fundamental conflict: the downsampling required for computational efficiency inevitably leads to the loss of features associated with small targets, thereby degrading detection performance. While one-stage detectors like the YOLO series offer speed and efficiency, they struggle to balance high detection accuracy with low model complexity, particularly in scenarios involving dense, small, and multi-scale objects common in aerial imagery. Existing solutions often force a trade-off where improving accuracy significantly increases parameters and computational load, or reducing parameters compromises model accuracy. Furthermore, deep convolutional networks often suffer from the attenuation of spatial positional information and feature mismatch between shallow and deep layers.

2. Methodology

To address these challenges, the authors propose RLPF-YOLO, a resolution-flexible, lightweight object detection model based on the YOLOv8n architecture. The design philosophy centers on "feature compensation," aiming to actively recover and enhance lost information through a closed-loop collaborative system of four key modules integrated across the backbone and neck:

2.1 Backbone Enhancements

• Dual-Branch Feature Complementary Module (C2F-FCM): Designed to mitigate spatial information attenuation, this module splits input features into a semantic-rich branch (1/4 channels) and a spatial processing branch (3/4 channels). It employs a cross-attention mechanism to interact these branches, implicitly encoding low-level spatial information into high-dimensional vectors. This ensures the integrity of feature representation in deeper layers by compensating for the lack of spatial details.
• Efficient Multi-Scale Convolution Module (RMKPConv): Replacing standard convolutions in the P5 layer, this module utilizes multi-scale depthwise convolutions (3×3, 5×5, 7×7) followed by pointwise convolutions. This approach captures contextual information across various scales with minimal computational cost, addressing the issue where large receptive fields in the P5 layer might overwhelm small targets with background noise.
• Residual Attention Module (RCBAM): Integrating residual connections with the Convolutional Block Attention Module (CBAM), this component dynamically emphasizes crucial spatial and channel features. It is specifically designed to enhance the semantic representation of shallow features and improve the network's adaptability to small objects and complex backgrounds.
• Down Convolution: Standard convolutions in the P4 and P5 layers are replaced with Down convolutions to optimize computational efficiency, alleviate the vanishing gradient problem, and stabilize feature updates while retaining small-target features.

2.2 Neck Enhancements

• Cross-Stage Partial Feature Processing Module (CSP-SimAM): In the neck section, the authors replace standard components with a combination of GSConv and CSP-SimAM. This module utilizes a cross-stage partial architecture enhanced with SimAM (a parameter-free attention mechanism). It facilitates efficient multi-scale fusion, removes computational redundancy, and reduces complexity without sacrificing accuracy.

3. Key Contributions

The paper outlines four primary contributions:

1. Spatial Information Recovery: The introduction of the C2F-FCM module addresses the degradation of spatial positional information in deep networks, enhancing model accuracy by ensuring consistent spatial semantics.
2. Small Object Semantic Enhancement: The RCBAM module, combined with residual connections, strengthens the semantic representation of shallow features, providing high-quality low-level features for subsequent network stages.
3. Optimized Downsampling: The implementation of RMKPConv in the P5 layer and the use of Down convolutions in P4/P5 layers significantly improve the retention rate of features for sparse-pixel small targets while optimizing computational efficiency.
4. Lightweight Neck Fusion: The integration of GSConv and CSP-SimAM in the neck achieves a balance between detection accuracy and reduced parameter count/computational complexity, making the model suitable for resource-constrained edge environments.

4. Experimental Results

Experiments were conducted on the VisDrone2019 dataset, which is characterized by dense small targets, occlusion, and varying scales.

• Performance at 640-pixel Resolution: Compared to the baseline YOLOv8n, RLPF-YOLO reduced the number of parameters by 70.4% (from 3.01M to 0.89M) and GFLOPS by 2.5%. Simultaneously, it increased mAP@50 from 32.2% to 34.2% and mAP@(50-95) from 18.6% to 20.1%.
• Performance at 1280-pixel Resolution: The model demonstrated excellent scale robustness, achieving an mAP@50 of 46.0% and mAP@(50-95) of 28.4%, outperforming the baseline in precision even at higher resolutions.
• Comparative Analysis: Against other state-of-the-art lightweight models (e.g., YOLOv7-tiny, YOLO-FBRT, YOLOv11n), RLPF-YOLO achieved the highest accuracy metrics (0.342 mAP@50) with the lowest parameter count (0.89M) among the compared high-performing models. For instance, compared to YOLOv7-tiny, it reduced parameters by 85.2% while improving accuracy.
• Ablation Studies: Step-by-step ablation experiments confirmed that each proposed module (C2F-FCM, Down Conv, RMKPConv, RCBAM, CSP-SimAM) contributed positively to the final performance, with the cumulative effect significantly outperforming the sum of individual improvements.

5. Significance and Claims

The authors claim that RLPF-YOLO provides an effective architectural solution to unify the typically opposing goals of lightweight design and high-accuracy small-object detection. By employing a "feature compensation" strategy, the model successfully recovers lost information caused by downsampling, enabling robust real-time detection on edge devices with limited computational resources. The work demonstrates that it is possible to achieve a significant reduction in model size and computational load while simultaneously improving detection accuracy for small and dense objects in complex UAV scenarios. The code is made available to support further research and deployment in resource-constrained environments.