Objective

2

Tens of accidents involving helicopters occur every year owing to collisions with trees

wires

poles

and man-made buildings at low altitude. Just in 2014—2016

there were 96 crashes caused by hitting power lines around the world. Thus

warnings and avoiding wires are important for the low-altitude flight safety of helicopters and unmanned aerial vehicles. According to relevant studies

utilization of optical images is an effective way to identify wires. Traditional methods use manual filters to extract features of power lines and then use Hough transform to detect the lines. Machine learning methods

such as VGG (visual geometry group) 16 and random forest (RF)

can only obtain a classification result for a picture

which makes confirming accuracy difficult. The full connection layer of the traditional convolutional neural network (CNN) is effective at classification tasks. However

it cannot carry out pixel segmentation tasks because of the loss of location information. By contrast

the fully convolutional network has no full connection layer

which misses location information. One kind of fully convolutional network

U-Net

is proposed to solve problems such as cell segmentation and retina segmentation. U-Net works well under the conditions of a small amount of samples and a small slice. A three-channel image is input into the network. Through the encoder and decoder

it finally becomes a one-channel feature map via 1×1 kernel size convolution. To obtain the final value between 0 and 1

Sigmoid activation function is used before every convolution layer. In this study

a CNN recognition method based on U-net is proposed to detect power lines.

Method

2

First

we obtain a power line data set containing 8 000 images with 4 000 pairs of visible and infrared images. The image size is 128×128 pixels

with each image having a pixel ground truth label. The network receptive field calculation formula is used to determine the depth of our network. Next

adjustments are made on this basis network to choose the best model. The basis network is named the U-Net-0 model. The U-Net-1 model removes the lower pooling layer in the U-Net-0 model and changes the step size of the convolution layer before the lower pooling layer to 2. It also removes the upper pooling layer and changes the convolution layer after the upper pooling layer to the inverse convolution layer with a step size of 2. Compared with U-Net-0

the U-Net-2 model eliminates the upper and lower pooling layers and the convolution layer in the middle

thereby reducing the network depth. In the U-Net-3 model

decoding is expected to be a dimensionality reduction process. Therefore

the number of convolution kernels of the decoding part is limited

and the number of parameters of feature graph output of each layer is not larger than that of the previous layer. Pictures with complex backgrounds are likewise used to generate a large number of paired synthetic data

including power line images with pixel labels. The generated synthetic data are then used for network training. For each image

the power line contains a small number of pixels. Thus

focal loss is used to balance the impact of a large number of negative samples. The four models use the same optimizer named "Adam"

which can automatically adjust the learning rate on the basis of SGD (stochastic gradient descent). The training procedure of each model is accelerated using an NVIDIA GTX 1080 TI device

which takes approximately 18 hours in 6 000 iterations with a batch size of 64. Loss

F1 score

and intersection-over-union (IoU) are the three evaluative criteria for trained models. The best model usually has low loss and high F1 score and IoU. Each model is used on visible and infrared images. The two results are combined to make a judgment. The power line

regardless of which of the same pair includes it

is finally considered detected in the mixed result.

Result

2

After these four models are tested on the data set

the number of correctly identified pixels and IoU on each image is counted. According to the statistical results

the IoU of most image recognition results exceeds 0.2

and the threshold of 30 pixels as the result classification is relatively good. If more than 30 pixels are identified on an image

this image might include a power line. By this standard

the proposed method achieves a recognition rate over 99%

while the false alarms are less than 2%. Moreover

VGG16

which is trained on 3 800 pairs of images and tested on 200 pairs of images

only obtains a recognition rate of 95% and a false alarm rate of 37%. RF is affected by feature extraction methods. Thus

the recognition rate and false alarm rate fluctuate greatly. For example

RF with local binary patterns has a recognition rate of 63.5% and a false alarm rate of 36.3% on infrared images. In addition

RF with discrete cosine transform obtains a recognition rate of 92.95% and a false alarm rate of 13.95% on infrared images. Although U-Net-3 has more learnable parameters than U-Net-2

its performance is substantially worse.

Conclusion

2

Our models have higher recognition rates and lower false alarm rates than do other traditional methods on the same dataset. Results show that our models are more effective than other methods and can even clearly extract power lines from background. Our models are trained on synthetic data and tested on real data

which means better generalization performance. The comparison of the four models also shows that the number of parameters cannot completely determine the performance of the network and that the reasonable structure is important. However

our current models have a small receptive field and cannot be used for power line recognition in high-resolution images. In the future

the models will be further studied to increase their receptive field for adapting to larger images without greatly increasing the number of parameters.