<p>As a core foundational function for robotic perception and execution, the accuracy and efficiency of grasp detection are critical to the overall performance of the system. However, current mainstream deep learning-based grasp detection methods often require large and deep network models to achieve the desired grasping accuracy, particularly in complex scenes or cluttered object piles. This reliance on extensive models results in high inference latency and significant computational resource consumption, presenting substantial challenges for lightweight model deployment. To address this issue, we propose a lightweight yet high-performance grasp detection network, LC-GraspNet. First, the model employs EfficientNetV2 as the backbone for extracting multi-level features . Subsequently, we introduce an adaptive multi-scale fusion module (LHAC Module) designed to dynamically integrate these multi-level features. This module utilizes a combination of dilated convolutions with varying rates and depthwise separable convolutions to efficiently model multi-scale features through dynamic parameter adjustment and multi-branch fusion, all while maintaining a lightweight design. Finally, adaptive decoder incorporate a self-attention to enhance global information modeling, thereby improving robustness in unstructured environments. The model has been rigorously trained and evaluated on the publicly available Cornell Grasp dataset and Jacquard dataset. The results indicate that the detection accuracy reached 99.6% and 97.13%. Compared to other methods in the literature, our approach achieves superior grasp detection accuracy, and the experimental results strongly validate the enhanced performance of the proposed model.</p>