3.2.4. Discussion

Five false targets are shown in Table 4. The first line shows that cable tower is detected as non-working chimney. The cable tower is highly similar to chimney in both texture feature and three-dimensional structure. The main direction of the image slice is 30.19◦, while the main direction of the whole image is 42.23◦. This difference may be caused by some decorative or structural curves on the cable tower, which makes it not so straight in the image. However, similar loaded or decorative component is seldom attached on a chimney, so the true chimney is unlikely to be mis-removed. In the second line, a big tank is mistakenly detected as a working condensing tower. They are similar in height, so cannot be distinguished by only introducing the DTM. However, its aspect ratio, which is much smaller than true condensing tower, make the calculation of main direction after binarization unstable, leading to a large different with the image main direction. For the chimney like objects (including condensing tower), which has large aspect ratio, the main direction is determined by the pixel value distribution of wall. For those with low aspect ratio, such as the oil tank, the main direction is highly affected by the pixel distribution of its top cover. Therefore, the main direction test is also useful to distinguish some objects with different aspect ratio. In line 3, a complex scene with working and non-working chimneys, oil tanks, and steam vents is shown. There are only two chimneys in this image, one undetected working chimney in the red circle. The reason why the working chimney in the red circle remains undetected is that the two spatial analysis methods introduced in this paper are ineffective to reduce the false negatives. We think that the improvement in detection ability of neural network and completeness of the training dataset might be helpful. The detected non-working chimney is in the upper left corner. The rest of the detected objects are all false. The objects with lower height, including a steam vent, can be removed by DTM filtering. The main direction test can remove all false target in line 3 because the main directions of most interfering targets are randomly distributed except some high vertical objects. However, it is possible that the main direction of interfering target is coincidently consistent with the main direction of the image. Two examples show in line 4 and 5. The false targets cannot be removed by main direction test are mainly ground texture, shadows or structure that caused by overlapping.

**Table 4.** Examples of four-class detection method results. The pink boxes represent working condensing tower, the green boxes represent non-working condensing tower, the blue boxes represent working chimney, and the yellow boxes represent non-working chimney.


The final evaluation indexes are shown in Table 5. The total target number (N) indicates the total chimneys in 3 images. The recall rates of four kinds of targets are 0.7727, 0.7662, 1, and 1, respectively. These values are much closed to the testing accuracies on BUAA-FFPP60 dataset. However, in practice, there is a large number of FPs, causing a very low precision. The original precisions are only 0.047, 0.4048, 0.2173, and 0.0833 for four kinds of target, respectively. After using two spatial analysis method, the FPs are largely removed. The precisions are increased to 0.9444, 0.9365, 0.833, and 0.8, respectively. The final qualities are 0.7391, 0.7108, 0.8333, and 0.8, respectively. The final qualities of working and nonworking chimneys are both significantly higher than the qualities calculated on testing samples. It can be concluded that the spatial analysis methods are very effective to increase the final precision and final quality.


**Table 5.** The accuracy of the experiment.

\* Three non-working chimneys are mis-removed.

In terms of category, chimneys have relatively low recall rate but high final precision. That is because the chimney is narrow in the image, and easily be interfered by noise, such as shadow, road, and build. Meanwhile, its unique contour makes it easy to distinguish with false chimney by spatial analysis method. In contrary, the condensing tower is easy to be detected by image-processing-based method, the Faster R-CNN, for its integrality appearance in image. Its relatively low final precision may partly result from the small number of samples.
