基于卷积神经网络的缺陷检测.pdf

发布时间：2022-06-19 发布人：admin 分类：说明书资料大小：6.15M 资料格式：pdf 举报版权申诉

17151f17-954f-42da-9705-e31ef99da8c8.pdf-第1页.png

第1页 / 共14页

17151f17-954f-42da-9705-e31ef99da8c8.pdf-第2页.png

第2页 / 共14页

17151f17-954f-42da-9705-e31ef99da8c8.pdf-第3页.png

第3页 / 共14页

17151f17-954f-42da-9705-e31ef99da8c8.pdf-第4页.png

第4页 / 共14页

17151f17-954f-42da-9705-e31ef99da8c8.pdf-第5页.png

第5页 / 共14页

17151f17-954f-42da-9705-e31ef99da8c8.pdf-第6页.png

第6页 / 共14页

17151f17-954f-42da-9705-e31ef99da8c8.pdf-第7页.png

第7页 / 共14页

17151f17-954f-42da-9705-e31ef99da8c8.pdf-第8页.png

第8页 / 共14页

文本预览

Demystifying Segnet Written by : Sirawat Pitaksarit Inner workings of Segnet is not very well documented. In this guide, I researched into Segnet by observations and script digging. Along with this, we can understand about neural network on implementation side like how people actually implement them. Demystifying Segnet First of all if you follow the previous guide Let’s look at SegNet’s output What does SegNet Tutorial’s training data looks like? What’s wrong with the CamVid training data Answering the questions Questions not answered Finally, how to train our own data? Conclusion 1 2 2 4 8 11 13 14 14 1

First of all if you follow the previous guide The demo’s Scripts/webcam_demo.py execution command is : python Scripts/webcam_demo.py --model Example_Models/ segnet_model_driving_webdemo.prototxt --weights Example_Models/ segnet_weights_driving_webdemo.caffemodel --colours Scripts/ camvid12.png The trained data is segnet_weights_driving_webdemo.caffemodel. This outputs colored images. Keep in mind that this image never contains the color black as we will come back to this later. Let’s look at SegNet’s output In Segnet-Tutorial repository, if you looked in Scripts/webcam_demo.py segmentation_ind = np.squeeze(net.blobs['argmax'].data) Is where we get the result from ﬁnal layer of the network.This result’s shape is (1,1,360,480) which is then np.squeeze-d into (360,480) I test the network that is supposed to segment road scene with the close up webcam picture of my face. Trying to print it and inspect the .shape, the output looks like this : [[ 1. 1. 1. ..., 0. 0. 0.] [ 1. 1. 1. ..., 0. 0. 0.] [ 1. 1. 1. ..., 0. 0. 0.] ..., [ 1. 1. 1. ..., 5. 5. 5.] [ 1. 1. 1. ..., 5. 5. 5.] [ 1. 1. 1. ..., 5. 5. 5.]]  (360, 480) 2

Each value corresponds to one pixel. The number of each values is the label. To see an example of how many instances of each label that exist in this output, we will try numpy.unique(A, False, False, True) to count each instant of them. (array([ 0., 1., 2., 4., 5., 6., 7., 9., 10., 11.], dtype=float32), array([51467, 19657, 297, 7994, 2120, 14383, 157, 70013, 6698, 14])) In this image there are many classes ranges from label 0 to 11. (3 and 8 is absent in this image) And the frequency of class label 0 is highest. The squeezed array then will be made into 3 channels and then used with lookup table function (np.LUT) to add colors. Now look at what SegNet can segment. You can see in the image : purple area is largest (Count : 70013) and that corresponds to vehicle (somehow my face looks like a vehicle), which is the label #9. Second largest is grey, label #0, sky. (Count : 51467) We can conclude that these numbers exactly aligns with horizontal list that Segnet web demo provided starting from #0 to #11. There is no vermillion color in my image, vermillion is label #3, “Road Marking” and the array debug shows no data that have value 3. (Skips from 2 to 4 as you 3

can see.) This proves that segnet_weights_driving_webdemo.caffemodel is indeed the same model as their web demo. One interesting point to think is the output from this never contains “Unknown”. This means segnet_weights_driving_webdemo.caffemodel has been trained with only known labels. To prove this, look at the Scripts/camvid12.png that we used to do lookup table color matching. Mouse is hovering at the ﬁnal light blue color. It is X=11. This means value 0~11 in the answer from network will get mapped to these colors. To get black color, it needs answer 12 and higher which never happen from this segnet_weights_driving_webdemo.caffemodel. All of these goes together well, but from next section I enountered more confusing things. What does SegNet Tutorial’s training data looks like? In this page, http://mi.eng.cam.ac.uk/projects/segnet/tutorial.html, the author shows how to train Segnet from scratch with provided CamVid data which you can get toghether with the Segnet- Tutorial repo. The author concludes that it will ends with 88.6% global test accuracy. You can follow the tutorial to learn how Caffe works on training/testing/evaluation. However, to train your own data or creating your own network, you need to understand how this works then you will be able to prepare your own data. In this section, I will prove that when ﬁnished training following the tutorial, one will not end up with the same model as segnet_weights_driving_webdemo.caffemodel (their web demo) that I tested in previous section. From previous section we can understand the input. Each training data to SegNet is a pair of image. The image data and the label. One such pair from Segnet-Tutorial/CamVid/train and trainannot looks like this 4

The second image is not black image. Each pixel is in greyscale format and each pixel has been labeled with number. This corresponds to the answer we got ealier, as the network directly uses these numbers as a target answer in the ﬁnal layer. The whole network will try to adjust weights and bias to produce these numbers as an answer for each pixels. If you repeatedly train with these almost black images you will get the network that can predict these small numbers. Then you can use the answer to color map to something meaningful. Let’s look at some numbers in this label image. As this number is very small, (for example R,G,B [0,0,0] [2,2,2] [3,3,3]) the result we see is nearly black. Using Photoshop’s Curve to increase the contrast between colors reveal some interesting data. In Photoshop’s “Info” panel, we can see the original value before using Curve. Red arrow is where the Info panel represented. (My screen capture program cannot capture mouse pointer) 5

We notice that something’s not according to our initial understanding. What’s wrong with the CamVid training data Let’s look at all the pictures earlier why these results looked strange. These area’s value is 14. But earlier we thought the network can only output number 0~11. 8

分享到：

赞收藏

资料库

基于卷积神经网络的缺陷检测.pdf

相关推荐

人工智能

热门标签

最新资料