Evaluation and Empirical Study of Image Enhancement Strategies for Small Object Detection in Remote Sensing

Optical remote sensing small object detection is limited by the low proportion of target pixels, complex backgrounds, and image degradation. To avoid equating visual enhancement with detection benefits, this paper analyzes degradation perception in image enhancement, combining image degradation, enhancement processing, target area visibility, and detection feedback, and empirically conducts the analysis based on the Northwestern Polytechnical University Very-High-Resolution 10-class (NWPU VHR-10) dataset. This paper combines the Common Objects in Context (COCO) small target area threshold and the relative area of the remote sensing target frame, and proposes a screening rule that the target frame area does not exceed 1% of the image area. Then, 30 positive samples were simulated for downsampling, blurring, noise, and haze degradation, comparing bicubic interpolation, Contrast Limited Adaptive Histogram Equalization (CLAHE), sharpening, dehazing, and fusion enhancement. After further training the You Only Look Once version 8 nano (YOLOv8n) detector, the mAP50 values for the original, degraded, and enhanced images were 0.921, 0.851, and 0.826, respectively, and the recall values were 0.871, 0.789, and 0.753, respectively. The results show that edge texture enhancement cannot be directly equated with detection benefits; enhancement evaluation should focus on full-image fidelity, target area indicators, and authentic detection feedback.