LDC: Lightweight Dense CNN for Edge Detection

Abstract

This paper presents a Lightweight Dense Convolutional (LDC) neural network for edge detection. The proposed model is an adaptation of two state-of-the-art approaches, but it requires less than 4% of parameters in comparison with these approaches. The proposed architecture generates thin edge maps and reaches the highest score (i.e., ODS) when compared with lightweight models (models with less than 1 million parameters), and reaches a similar performance when compare with heavy architectures (models with about 35 million parameters). Both quantitative and qualitative results and comparisons with stateof-the-art models, using different edge detection datasets, are provided. The proposed LDC does not use pre-trained weights and requires straightforward hyper-parameter settings. The source code is released at https://github.com/xavysp/LDC.

Dense Extreme Inception Network: Towards a Robust CNN Model for Edge Detection

“…Based on the quantitative results from DexiNed [41] and LDC [39], we trained all models with BIPED. Table 2 presents results from TEED and the state-of-the-art models with less than 1M parameters; all these models are trained with BIPED and evaluated with UDED.…”

“…These advantages are plausible in many computer vision tasks. Following these successes, DexiNed [40] and LDC [39] use similar architectures for edge detection. The result is a model trained from scratch that still achieves state-of-the-art accuracy.…”

“…This skip-connection is applied immediately after the max-pooling operation. Another skip connection, skip2, is used to fuse h 2 and the suboutputs of block 3 as follow (Block3.1+h2)/2; this is similar to the skip connection in [40] and [39]. The outputs of these three blocks feed USNet, as indicated by gray arrows pointing downwards in Fig.…”

“…It contains 250 images in high definition (720 × 1280): 50 images of which were selected by the authors for testing and the rest for training and validation. The data augmentation procedure used in LDC [39] is implemented for TEED.…”

“…Validation results at epoch 6 are reported. Despite the fact that BIPED provides binary annotations, due to the interpolations when edge maps are built floating point values appear; hence we apply a transformation to edge annotations y by adding 0.2 to values greater than 0.1, and then clipping the results to [0,1], similar to LDC [39]. A Lenovo Yoga C740-15IML laptop with an Intel i5-10210U processor is used to report the FPS values.…”