Target detection based on improved YOLOv7-tiny

As deep learning technologies continually advance, enhanced performance metrics are being realized in the realm of object detection. To address the challenge of ship detection in optically captured imagery set against intricate terrestrial and maritime backdrops, this study introduces a refined detection paradigm founded on YOLOv7-tiny. Initially, the backbone network juncture of the baseline model is supplanted by the streamlined MobileNetV3 architecture. Subsequently, a weighted Bi-directional Feature Pyramid Network (BiFPN) is integrated at the feature amalgamation stage to facilitate cross-scale coherence and weighted attribute fusion. Lastly, anchor frames optimization is achieved via the K-Means technique, and the CIOU loss function is supplanted by the EIOU loss. Evaluated on the Seaships (7000) dataset, the augmented model exhibits a detection precision of 98.3%, a enhancement of 0.2% relative to its predecessor. Concurrently, the computational and parametric overheads are curtailed to 6.9GFLOPs and 4.18M, representing reductions of 47% and 31%, respectively. Thus, the optimized YOLOv7-tiny framework offers a computationally efficient yet highly accurate solution for ship detection tasks.