基于梯度和模糊监督的边缘检测OA

Currently,edge detection models have reached the level of human perception of edges,but models with excellent perception have a larger number of parameters and slower inference speeds,making them difficult to meet the needs in practical scenarios.Lightweight models have characteristics of small parameter sizes and fast inference speeds,but due to the limitation of model parameter numbers,they lag behind the afore-mentioned models in the comprehensive performance of precision and recall.This paper notes that existing methods use supervised signals in an insufficient and unreasonable manner during model training and accordingly proposes a training strategy of gradient and blur supervision.In training,the gradient information from labeled images is added as additional supervised signals to achieve more sufficient use of supervised signal,and the operation of blurring deep supervised signals aims to achieve a more reasonable supervision for different stages of the model.This paper applies the proposed method to models and achieve further improvements in the comprehensive performance of precision and recall of lightweight model without degrading model performance.