3.2.2. The Results from Faster R-CNN

After, we get the trained model. The Google images were input to the trained Faster R-CNN network. Due to the large area, the entire image is detected by window. The window size is 700 × 700 pixels and the step length is 500 pixels. The overlapped area in each step is as wide as 200 pixels, which is wide enough to prevent missing detection of chimneys at the edge of image. In order to detect more targets, we add an image enhancement method by adjusting the brightness and contrast ratio before Faster R-CNN detection. We also set a low network detection probability threshold, which is 0.3, to reduce the false negative and increase the recall rate.

In order to analyze the detection accuracy, we divide the detection results into nine types: working chimneys, non-working chimneys, working condensing towers, non-working condensing towers, road, architecture, tank, lake, topography. Figure 5 shows some examples of false detection.

**Figure 5.** The false detection objects are divided into five categories: lake, road, architecture, tank, and other objects. The pink boxes represent working condensing tower, the green boxes represent non-working condensing tower, and the yellow boxes represent non-working chimney.

It can be found from Table 2 that the road and architecture are most likely to be mis-detected as chimneys, the number of which are 45 and 59 respectively. Condensing towers are most likely to be mixed up by tanks and lakes. The false detection rate of working chimneys, non-working chimneys, working condensing towers and non-working condensing towers are 0.5952, 0.5810, 0.8214, and 0.9166, respectively.


**Table 2.** Faster R-CNN detection result.

3.2.3. The Results from Faster R-CNN, Elevation Filtering, and Main Direction Test

It can be found from Table 3 that by using the detection and test method—most of the false chimneys are removed. The false detection rate of working chimneys, non-working chimneys, working condensing towers and non-working condensing towers are significantly reduced to 0.0555, 0.0634, 0.1667, and 0.2, respectively. Meanwhile, only three non-working chimneys are mis-removed. That means after processing the true chimneys have been well retained.


**Table 3.** Faster R-CNN+ elevation filtering + main direction detection result.
