Edge detector based on multi-scale feature fusion: a robust convolutional neural network–based edge detector

Exploiting multi-scale representations is crucial to improve edge detection of objects at different scales. However, related research often ignores that edge information mainly exists in shallow features and blindly increases the extraction of deep semantic information without paying attention to shallow information. This makes it difficult for deep models to capture sufficient fine-grained details from these sparse shallow features. To address this issue, we propose an edge detector based on multi-scale feature fusion (EDMFF), a lightweight convolutional neural network–based edge detector. It adopts a shallow feature weighted extraction strategy to enrich the edge information during learning, thereby providing fine-grained features critical for multi-scale fusion. Building on this, we design a fused feature extraction module (Pfusion) based on partial convolution, which selectively convolves specific channels to achieve precise multi-dimensional feature fusion. Experimental results on the edge detection generalization dataset show that EDMFF exhibits faster inference speed among similar models under the 10M parameter scale. Compared with the state-of-the-art model, it achieves improvements in global optimal threshold, single image optimal threshold, average precision, and precision at 50% recall by +0.003, +0.013, +0.091, and +0.14, respectively. Its practical value is further validated by superior accuracy in downstream tasks such as sketch image retrieval.