Скачать с ютуб HSV Color Range Thresholding - OpenCV Object Detection in Games #6 в хорошем качестве

HSV Color Range Thresholding - OpenCV Object Detection in Games #6

Improve your object detection by using the HSV Thresholding technique in OpenCV. I'll also show you how to use the OpenCV GUI builder to adjust your image processing in real-time. This is a Python tutorial. Full tutorial playlist:    • OpenCV Object Detection in Games   Grab the code on GitHub: https://github.com/learncodebygaming/... 0:58 Refactoring find() 4:08 HSV thresholding theory 5:34 Building an OpenCV GUI for filter controls 7:30 Creating a custom data structure in Python 8:49 Applying an HSV color filter 11:02 Increasing saturation and brightness of an image 12:43 Color filter demonstration 16:36 Combining HSV filtering with object detection 20:33 Few more tweaks for find() Read the full written tutorial with code samples here: https://learncodebygaming.com/blog/hs... In the previous steps, we learned how to apply OpenCV's matchTemplate() to a video game in real-time. This code achieves object detection, but in many situations it's not working as well as we'd like. We can get better results by first processing our images before sending them through to matchTemplate(). In this article, we'll discuss one such method for pre-processing: Hue-Saturation-Value filtering (HSV range thresholding). In preparation to potentially apply multiple types of filters in the future, let's begin by refactoring our find() method in the Vision class. We'll break this up into different methods for greater flexibility. Let's make find() simply return the rectangle results from match templates. I've added additional code here to limit the maximum number of results returned, as well as to return an empty list that can easily be joined with result lists from other find() calls. Next we'll write get_click_points(self, rectangles), which will return a list of [x, y] coordinates at the midpoint of each rectangle. This is the same code we had in find() before, but now it's untangled from the debug output. Finally, let's write separate methods for draw_rectangles() and draw_crosshairs(). These will replace our debug output from before. And instead of having these methods be responsible for calling cv.imshow(), let's instead simply return the output images and move the cv.imshow() responsibility to our main file. In our main loop, we can now call find() to get the rectangle results from matchTemplate(), call draw_rectangles() to get the screenshot image with those rectangles drawn on it, and then give that processed image to cv.imshow(). Now's a good time to test and confirm everything is still working as before. With our vision code set up like this, we now have the flexibility to mix and match what sort of processing we do on each image before ultimately displaying it. And we can do that processing either before or after calling find() to search for our needle image. Our goal in the rest of this series is to use additional processing to make our object detection targets easier for matchTemplate() to recognize. We'll begin this exploration with HSV color filtering. Up to this point, we've been working with images in BGR format (Blue-Green-Red). In this format, each of these three color channels is represented by a number from 0 to 255, where 0 is the complete lack of this color and 255 is the complete presence of this color. If the object we are looking to detect is blue, we might try ignoring the green and red channels and just focus on the blue, where our object really pops out. This could be done by changing the green and red value on every pixel in our image to 0, so that only the blue values remain. This would eliminate any non-blue objects from possibly causing a false detection. We could go one step further if our object was always bright blue, and we wanted to ignore dark blue objects. In that case, we could look at the blue value of every pixel and if it's below a certain value (say 150) change that value to 0. This would effectively eliminate all dark blue objects. The technique we've just described is called color filtering, or range thresholding (in this context they mean the same thing). By pre-processing an image in this way, it reduces the chance of false-positives, which in turn allows us to lower the match threshold we give matchTemplate. The net result is we can detect a wider variety of objects from a single template image, without also exploding our rate of incorrect detections. Continue with the written tutorial here: https://learncodebygaming.com/blog/hs...