LMED: A Lightweight Multi-Scale Edge Detection Model for Crushed Rock Slag

In the filling of highway subgrade, the size of qualified crushed rock slag is one of the key factors for project quality. This paper processes the surface images, extracts the edges of the crushed rock slag in the images. Existing edge detection methods often face difficulties in balancing accuracy and computational efficiency, particularly when handling complex scenes. To address these challenges, this paper presents LMED: a lightweight edge detection model that combines multi-scale feature extraction with dilated convolutions. The model employs the USNet module, which uses convolutions with different dilation rates to extract multi-scale edge features. Simultaneously, the ResDilatedBlock module leverages dilated convolutions to expand the receptive field and integrates residual learning to enhance contextual information while maintaining efficient computation. To optimize model performance, a weighted hybrid loss function is designed to guide the training process of the deep learning model. When evaluated under the same conditions, our model uses only 80% of the parameters of similar models while the ODS value is nearly twice as high, demonstrating the superiority of our model in lightweight and high-precision detection. Moreover, the model's transferability to specific tasks, such as particle segmentation, is also validated.