Lightweight multidimensional feature enhancement algorithm LPS-YOLO for UAV remote sensing target detection

Abstract

Detecting small targets in UAV remote sensing images is challenging for traditional lightweight methods due to difficulty in feature extraction and high background interference. We propose LPS-YOLO, which improves small target feature extraction while reducing computational complexity by replacing the Conv backbone with SPDConv to retain fine-grained features. LPS-YOLO introduces the SKAPP module for better feature fusion and incorporates the E-BiFPN and OFTP structures to efficiently preserve and transfer backbone information. Evaluation of the VisDrone2019 dataset shows a 17.3% increase in mean Average Precision (mAP) and a 42.5% reduction in parameters compared to the baseline. Additional experiments on the DOTAv2 dataset demonstrate the model's robustness, with a 14.5% improvement in F1 score and a 14.9% increase in mAP over YOLOv8-n. LPS-YOLO offers an effective solution for multi-target detection in UAVs.

Fig. 1

Illustration of SPD-Conv when scale=2.

Fig. 2

Overall structure of the SKAPP module. Notice that $\otimes$ represents Hadamard product, k represents the maximum receptive field, and r represents the dilation rates.

Fig. 3

Four CBS modules are inserted between the backbone and neck to store feature information from top to bottom. For ease of presentation, we designate the A1 layer in the backbone section as the integration module comprising CBS, SPDConv, and C2f. Sequentially, layers A2, A3, and A4 are defined as integration modules combining SPDConv and C2f. Layer A5 is specified as the SKAPP module. Furthermore, layers B1, B2, B3, C1, and C2 within the neck and head sections are categorized as integrated modules of E-BiFPN and C2f.

Fig. 4

There are two structures of E-BIPFN in the model. The dotted line indicates a repeat block.

Fig. 5

Design of LPS-YOLO. The purple-filled modules are the ones that were modified.

Fig. 6

(a) Category distribution of the DOTAv2 dataset. (b) Category distribution of the VisDrone2019 dataset. Some categories in the figure are represented by abbreviations: Person, Van, etc. for small-sized targets; Plane, Airport, etc. for large-sized targets.

Fig. 7

(a) Comparison of each model between the number of parameters and mAP@0.5. (b) Comparison of FPS and mAP@0.5 across models.

Fig. 8

Visualization results of challenging images on the VisDrone2019 testing dataset. In the picture, different categories are represented by boxes with different colors, and the numbers on the rectangles represent scores with a confidence level above 0.25.

Fig. 9

We use feature heat maps to represent each layer of modules. The stronger the feature fusion ability, the clearer the shape outline of the graph. (a) The characteristic heat map of YOLOv8, (b) The characteristic heat map of LPS-YOLO.