YOLOv9-QX: a feature fusion and enhancement-based algorithm for underwater object detection

Abstract To address the issues of target blurriness, complex underwater backgrounds, target occlusion, and low detection accuracy for small targets in underwater object detection, this paper proposes an improved YOLOv9-QX object detection model. First, to tackle the issue of target blurriness, a Feature mixed convolution module (FMC) is introduced, which integrates global features and enhances the global perceptual capability. Second, to address the challenges posed by complex underwater backgrounds and target occlusion, a multiscale fusion module is proposed. This module is strategically positioned at the interface between the backbone network and the feature extraction network in YOLOv9, facilitating the effective fusion of features from both networks. Finally, to improve the detection accuracy of small targets, a feature enhanced aggregation module (FEA) is proposed. The FEA employs upsampling and feature fusion techniques to refine image details, enhancing the distinctiveness of small target features and improving their identifiability. The experimental results on the DUO dataset demonstrate that the proposed algorithm achieves a detection accuracy of 88.08%, representing a 1.58% improvement over the baseline model. Additionally, the detection speed reaches 163.3 frames per second, achieving an optimal balance between detection accuracy and computational efficiency.