Current steel surface defect detection methods have the shortcomings of low accuracy, poor real-time performance and limitation to detect small targets. To solve those problems, we proposed a lightweight detection model named GMD-YOLOv5s. Firstly, we introduce the C3 Global Ghost (C3-GGhost) module, an enhancement of the original C3 module, which improves both detection speed and accuracy for industrial applications. Secondly, Multi-Scale Dilated Attention (MSDA) is integrated into the neck network to capture target at different scales. Because MSDA can expand the model's receptive field and feature representation capabilities. Finally, in order to enhance the model's ability to detect small targets, we replace the conventional detection head with a Distributed Pixel-level Feature Refinement Head (DPFRHead). Experimental results show that GMD-YOLOv5s achieves a mean Average Precision (mAP) of 82.9% on the NEU-DET dataset. Our method improves by 30% compared to the baseline YOLOv5, while reducing the number of parameters by 60%. Ablation and comparative experiments validate the model's effectiveness and accuracy in detecting defects on steel surfaces.

Effective weed management is essential for maintaining stable agricultural productivity and ensuring crop yield. Recent advancements in computer vision technologies have significantly enhanced weed detection efficiency. However, the diversity of weed species and irregular spatial distribution present considerable challenges to traditional weed detection methods, which often exhibit low accuracy in practical applications. Additionally, the construction of large-scale, high-quality datasets encompassing a wide variety of weed species is constrained by high costs and time limitations. Indirect weed recognition methods offer a promising solution to this issue. This approach first employs an image segmentation network to precisely identify and remove crop regions from the original image, followed by an image processing algorithm that extracts green pixels for weed identification. Based on the indirect weed detection method, this study optimizes the YOLOv8n-seg architecture to improve segmentation accuracy while maintaining model efficiency, leading to the development of the enhanced ESL-YOLOv8 segmentation model. To validate the effectiveness of the proposed method, a maize-specific segmentation dataset was constructed, and extensive experiments were conducted. The results show that the improved ESL-YOLOv8 model achieves a mask mAP50 value of 94.1% based on the mask, which is superior to the original YOLOv8nseg(92.0%). Furthermore, the model size is reduced to 5.7 MB, compared to 6.8 MB for the original model, with a decrease in computational complexity. These findings confirm that the proposed framework provides a lightweight, high-precision, and practical solution for weed detection in agricultural fields, exhibiting enhanced robustness and significant engineering applicability.

Underwater small object detection is vital for marine monitoring, archaeology, and resource exploration. However, complex backgrounds and small target sizes present significant challenges. Traditional methods often lack the feature extraction accuracy and efficiency required for practical, lightweight applications. Although YOLOv8 offers high accuracy, its backbone and feature extraction modules are computationally intensive, limiting deployment on resource-constrained devices. To address this, we propose a lightweight detection framework based on YOLOv8 with three key improvements: (1) replacing the original backbone with StarNet to enhance efficiency and reduce computation; (2) introducing Wavelet Transform Convolution (WTConv) in the C2f module to expand the receptive field and reduce redundancy; (3) incorporating a Convolutional Additive Self-Attention (CAS) mechanism in the Neck to improve feature fusion. Experiments on the URPC dataset show that our model achieves 85.8% mAP, surpassing mainstream methods (84.7%) while reducing parameters (2.2M vs. 2.4M) and computation (5.8G vs. 7.5G), demonstrating superior accuracy and lightweight design.

Underwater target detection presents substantial technical challenges due to the complex and dynamic characteristics of marine environments. Detection performance is frequently impaired by factors such as the presence of small targets with limited visual features, frequent occlusions from marine organisms and suspended particles, as well as significant morphological variations caused by varying viewing angles, lighting conditions, and target deformations in aquatic environments. To tackle these challenges, this paper presents an improved YOLOv8-based model, referred to as FCD-YOLO. The proposed model incoporates a FCA mechanism to emphasize key features while suppressing background noise. Additionally, it integrates a CCFF Module to effectively capture and merge multi-scale information for more comprehensive feature representation. To further improve adaptability, the original YOLOv8 detection head is replaced with a DyHead, enabling multi-scale feature detection and enhancing the model’s ability to handle targets of varying sizes. Experimental results on the URPC2020 dataset demonstrate that the FCD-YOLO model achieves a mean Average Precision (mAP) of 86.9%, representing a 5.0% improvement over the baseline YOLOv8 model. Comparative studies with other state-of-the-art detectors further validate FCD-YOLO’s superior accuracy. Ablation studies confirm the individual contributions of each component, with the FCA mechanism and DyHead module showing particularly significant impacts on small target detection performance. These results substantiate FCD-YOLO’s superiority and practical utility in complex underwater environments.