LS-Net: fast single-shot line-segment detector

Abstract

In unmanned aerial vehicle (UAV) flights, power lines are considered as one of the most threatening hazards and one of the most difficult obstacles to avoid. In recent years, many vision-based techniques have been proposed to detect power lines to facilitate self-driving UAVs and automatic obstacle avoidance. However, most of the proposed methods are typically based on a common three-step approach: (i) edge detection, (ii) the Hough transform, and (iii) spurious line elimination based on power line constrains. These approaches not only are slow and inaccurate but also require a huge amount of effort in post-processing to distinguish between power lines and spurious lines. In this paper, we introduce LS-Net, a fast single-shot line-segment detector, and apply it to power line detection. The LS-Net is by design fully convolutional, and it consists of three modules: (i) a fully convolutional feature extractor, (ii) a classifier, and (iii) a line segment regressor. Due to the unavailability of large datasets with annotations of power lines, we render synthetic images of power lines using the physically based rendering approach and propose a series of effective data augmentation techniques to generate more training data. With a customized version of the VGG-16 network as the backbone, the proposed approach outperforms existing state-of-the-art approaches. In addition, the LS-Net can detect power lines in near real time. This suggests that our proposed approach has a promising role in automatic obstacle avoidance and as a valuable component of self-driving UAVs, especially for automatic autonomous power line inspection.

A Novel Focal Phi Loss for Power Line Segmentation with Auxiliary Classifier U-Net

“…To the best of authors’ knowledge, almost all the studies regarding PL detection use the dice scores (DSC) (also known as the F1-score), precision, true positive rate (TPR) (also known as recall or sensitivity), false discovery rate (FDR) and accuracy [ 4 , 6 , 7 , 8 , 12 , 13 , 14 ]. These evaluation parameters are defined as: DSC or F1-score = 2TP/(2TP + FP +FN) Precision = TP/(TP + FP) TPR or Recall or Sensitivity = TP/(TP + FN) FDR = FP/(FP + TP) Accuracy = (TP + TN)/(TP + TN + FP + FN) where TP, TN, FP and FN represent the true positive, true negative, false positive and false negative entries of the confusion matrix, respectively.…”

“…To the best of our knowledge, no research work has investigated the application of different loss functions for the highly imbalanced PL detection task with thin RoIs. All the studies on PL detection mainly rely on BCE loss [ 6 , 14 ] and its variants [ 4 , 7 , 12 ] to handle the class imbalance problem, with the BBCE loss being the baseline in the majority of these works. The same ACU-Net network architecture is trained with each of these loss functions and the characteristic evaluation parameters are monitored.…”

“…To the best of authors’ knowledge, all the studies on PL detection utilize the BCE loss [ 6 , 14 , 15 , 16 ] and its class imbalance variants [ 4 , 7 , 12 ] for segmenting the PLs. Although BCE loss is easier to optimize with lower training times, it might not always be the best choice for training deep classification networks [ 17 , 18 ].…”

“…The loss functions shift the focus to the minority class, thus, allowing it to contribute more to the loss in comparison to the major background class. Despite the fact that class imbalance affects the PL detection task adversely, only a handful of PL detection works address the class imbalance problem either via the popular variant of the binary cross entropy (BCE) loss function, namely the balanced BCE (BBCE) loss [ 4 , 7 , 12 ], or via a compound loss function [ 8 ] or the attention gated networks [ 13 ]. However, other notable studies on PL detection do not consider the prevalent class imbalance problem and utilize the vanilla BCE loss function for training the deep semantic segmentation networks [ 6 , 14 , 15 , 16 ].…”

“…However, other notable studies on PL detection do not consider the prevalent class imbalance problem and utilize the vanilla BCE loss function for training the deep semantic segmentation networks [ 6 , 14 , 15 , 16 ]. Moreover, BCE loss and its variants are also the usual choice for loss function in the PL detection tasks [ 4 , 6 , 7 , 12 , 14 , 15 , 16 ]. Nevertheless, it has been demonstrated that other kinds of losses can outperform the popular BCE loss for typical classification tasks [ 17 , 18 ].…”