Modern GUI Applications for Computer Vision in Python

I’m an enormous fan of interactive visualizations. As a pc imaginative and prescient engineer, I deal virtually every day with picture processing associated duties and most of the time I’m iterating on an issue the place I would like visible suggestions to make choices. Let’s consider a quite simple picture processing pipeline with a single step that has some parameters to rework a picture:

How have you learnt which parameters to regulate? Does the pipeline even work as anticipated? With out visualizing your output, you may miss out on some key insights and make sub optimum selections.

Generally merely exhibiting the output picture and/or some calculated metrics may be sufficient to iterate on the parameters. However I’ve discovered myself in lots of conditions the place a instrument could be immensely useful to iterate shortly and interactively on my pipeline. So on this article I’ll present you work with easy built-in interactive parts from OpenCV in addition to construct extra trendy consumer interfaces for Laptop Imaginative and prescient initiatives utilizing customtkinter.

Stipulations

If you wish to observe alongside, I like to recommend you to arrange your native setting with uv and set up the next packages:

uv add numpy opencv-Python pillow customtkinter

Aim

Earlier than we dive into the code of the venture, let’s shortly define what we need to construct. The applying ought to use the webcam feed and permit the consumer to pick out various kinds of filters that can be utilized to the stream. The processed picture needs to be proven in real-time within the window. A tough sketch of a possible UI would look as follows:

OpenCV – GUI

Let’s begin with a easy loop that fetches frames out of your webcam and shows them in an OpenCV window.

import cv2

cap = cv2.VideoCapture(0)

whereas True:
    ret, body = cap.learn()
    if not ret:
        break

    cv2.imshow("Video Feed", body)
    
    key = cv2.waitKey(1) & 0xFF
    if key == ord('q'):
        break

cap.launch()
cv2.destroyAllWindows()

Keyboard Enter

The only manner so as to add interactivity right here, is by including keyboard inputs. For instance, we will cycle by means of completely different filters with the quantity keys.

...

filter_type = "regular"

whereas True:
    ...

    if filter_type == "grayscale":
        body = cv2.cvtColor(body, cv2.COLOR_BGR2GRAY)
    elif filter_type == "regular":
        move

    ...

    if key == ord('1'):
        filter_type = "regular"
    if key == ord('2'):
        filter_type = "grayscale"
        
    ...

Now you possibly can swap between the traditional picture and the grayscale model by urgent the quantity keys 1 and a pair of. Let’s additionally shortly add a caption to the picture so we will truly see the identify of the filter we’re making use of.

Now we must be cautious right here: if you happen to check out the form of the body after the filter, you’ll discover that the dimensionality of the body array has modified. Keep in mind that OpenCV picture arrays are ordered HWC (peak, width, colour) with colour as BGR (inexperienced, blue, pink), so the 640×480 picture from my webcam has form (480, 640, 3).

print(filter_type, body.form)
# regular (480, 640, 3)
# grayscale (480, 640)

Now as a result of the grayscale operation outputs a single channel picture, the colour dimension is dropped. If we now need to draw on prime of this picture, we both have to specify a single channel colour for the grayscale picture or we convert that picture again to the unique BGR format. The second choice is a bit cleaner as a result of we will unify the annotation of the picture.

if filter_type == "grayscale":
    body = cv2.cvtColor(body, cv2.COLOR_BGR2GRAY)
elif filter_type == "regular":
    move

if len(body.form) == 2: # Convert grayscale to BGR
    body = cv2.cvtColor(body, cv2.COLOR_GRAY2BGR)

Caption

I need to add a black border on the backside of the picture, on prime of which the identify of the filter can be proven. We are able to make use of the copyMakeBorder perform to pad the picture with a border colour on the backside. Then we will add the textual content on prime of this border.

# Add a black border on the backside of the body
border_height = 50
border_color = (0, 0, 0)
body = cv2.copyMakeBorder(body, 0, border_height, 0, 0, cv2.BORDER_CONSTANT, worth=border_color)

# Present the filter identify
cv2.putText(
    body,
    filter_type,
    (body.form[1] // 2 - 50, body.form[0] - border_height // 2 + 10),
    cv2.FONT_HERSHEY_SIMPLEX,
    1,
    (255, 255, 255),
    2,
    cv2.LINE_AA,
)

That is how the output ought to look, and you’ll swap between the traditional and grayscale mode and the frames can be captioned accordingly.

Sliders

Now as an alternative of utilizing the keyboard as enter methodology, OpenCV affords a fundamental trackbar slider UI ingredient. The trackbar must be initialized originally of the script. We have to reference the identical window as we can be exhibiting our photographs in later, so I’ll create a variable for the identify of the window. Utilizing this identify, we will create the trackbar and let it’s a selector for the index within the record of filters.

filter_types = ["normal", "grayscale"]

win_name = "Webcam Stream"
cv2.namedWindow(win_name)

tb_filter = "Filter"
# def createTrackbar(trackbarName: str, windowName: str, worth: int, depend: int, onChange: _typing.Callable[[int], None]) -> None: ...
cv2.createTrackbar(
    tb_filter,
    win_name,
    0,
    len(filter_types) - 1,
    lambda _: None,
)

Discover how we use an empty lambda for the onChange callback, we are going to fetch the worth manually within the loop. Every part else will keep the identical.

whereas True:
    ...

    # Get the chosen filter sort
    filter_id = cv2.getTrackbarPos(tb_filter, win_name)
    filter_type = filter_types[filter_id]

    ...

And voilà, we have now a trackbar to pick out our filter.

Now we will additionally simply add extra filters simply by extending our record and implementing every processing step.

filter_types = [
    "normal",
    "grayscale",
    "blur",
    "threshold",
    "canny",
    "sobel",
    "laplacian",
]

...

    if filter_type == "grayscale":
        body = cv2.cvtColor(body, cv2.COLOR_BGR2GRAY)
    elif filter_type == "blur":
        body = cv2.GaussianBlur(body, ksize=(15, 15), sigmaX=0)
    elif filter_type == "threshold":
        grey = cv2.cvtColor(body, cv2.COLOR_BGR2GRAY)
        _, thresholded_frame = cv2.threshold(grey, thresh=127, maxval=255, sort=cv2.THRESH_BINARY)
    elif filter_type == "canny":
        body = cv2.Canny(body, threshold1=100, threshold2=200)
    elif filter_type == "sobel":
        body = cv2.Sobel(body, ddepth=cv2.CV_64F, dx=1, dy=0, ksize=5)
    elif filter_type == "laplacian":
        body = cv2.Laplacian(body, ddepth=cv2.CV_64F)
    elif filter_type == "regular":
        move

    if body.dtype != np.uint8:
        # Scale the body to uint8 if obligatory
        cv2.normalize(body, body, 0, 255, cv2.NORM_MINMAX)
        body = body.astype(np.uint8)

Trendy GUI with CustomTkinter

Now I don’t learn about you however the present consumer interface doesn’t look very trendy to me. Don’t get me improper, there’s some magnificence within the fashion of the interface, however I desire cleaner, extra trendy designs. Plus we’re already on the restrict of what OpenCV affords out of the field by way of UI parts. Yep, no buttons, textual content fields, dropdowns, checkboxes or radio buttons and no customized layouts. So let’s see how we will remodel the look and consumer expertise of this fundamental software to a recent and clear one.

So to get began, we first have to create a category for our app. We create two frames: the primary one comprises our filter choice on the left aspect and the second wraps the picture show. For now, let’s begin with a easy placeholder textual content. Sadly there’s no out of the field opencv part from customtkinter straight, so we might want to shortly construct our personal within the subsequent few steps. However let’s first end the fundamental UI structure.

import customtkinter


class App(customtkinter.CTk):
    def __init__(self) -> None:
        tremendous().__init__()

        self.title("Webcam Stream")
        self.geometry("800x600")

        self.filter_var = customtkinter.IntVar(worth=0)

        # Body for filters
        self.filters_frame = customtkinter.CTkFrame(self)
        self.filters_frame.pack(aspect="left", fill="each", increase=False, padx=10, pady=10)

        # Body for picture show
        self.image_frame = customtkinter.CTkFrame(self)
        self.image_frame.pack(aspect="proper", fill="each", increase=True, padx=10, pady=10)

        self.image_display = customtkinter.CTkLabel(self.image_frame, textual content="Loading...")
        self.image_display.pack(fill="each", increase=True, padx=10, pady=10)

app = App()
app.mainloop()

Filter Radio Buttons

Now that the skeleton is constructed, we will begin filling in our elements. For the left aspect, I can be utilizing the identical record of filter_types to populate a gaggle of radio buttons to pick out the filter.

        # Create radio buttons for every filter sort
        self.filter_var = customtkinter.IntVar(worth=0)
        for filter_id, filter_type in enumerate(filter_types):
            rb_filter = customtkinter.CTkRadioButton(
                self.filters_frame,
                textual content=filter_type.capitalize(),
                variable=self.filter_var,
                worth=filter_id,
            )
            rb_filter.pack(padx=10, pady=10)

            if filter_id == 0:
                rb_filter.choose()

Picture Show Element

Now we will get began on the fascinating half, get our OpenCV frames to indicate up within the picture part. As a result of there’s no built-in part, let’s create our personal primarily based on the CTKLabel. This permits us to show a loading textual content whereas the webcam stream is beginning up.

...

class CTkImageDisplay(customtkinter.CTkLabel):
    """
    A reusable ctk widget widget to show opencv photographs.
    """

    def __init__(
        self,
        grasp: Any,
    ) -> None:
        self._textvariable = customtkinter.StringVar(grasp, "Loading...")
        tremendous().__init__(
            grasp,
            textvariable=self._textvariable,
            picture=None,
        )

...

class App(customtkinter.CTk):
    def __init__(self) -> None:
        ...

        self.image_display = CTkImageDisplay(self.image_frame)
        self.image_display.pack(fill="each", increase=True, padx=10, pady=10) 

To date nothing has modified, we merely swapped out the present label with our customized class implementation. In our CTKImageDisplay class we will outline an perform to indicate a picture within the part, let’s name it set_frame.

import cv2
import numpy.typing as npt
from PIL import Picture

class CTkImageDisplay(customtkinter.CTkLabel):
    ...

    def set_frame(self, body: npt.NDArray) -> None:
        """
        Set the body to be displayed within the widget.

        Args:
            body: The brand new body to show, in opencv format (BGR).
        """
        target_width, target_height = body.form[1], body.form[0]

        # Convert the body to PIL Picture format
        frame_rgb = cv2.cvtColor(body, cv2.COLOR_BGR2RGB)
        frame_pil = Picture.fromarray(frame_rgb, "RGB")

        ctk_image = customtkinter.CTkImage(
            light_image=frame_pil,
            dark_image=frame_pil,
            dimension=(target_width, target_height),
        )
        self.configure(picture=ctk_image, textual content="")
        self._textvariable.set("")

Let’s digest this. First we have to understand how huge our picture part can be, we will extract that info from the form property of our picture array. To show the picture in tkinter, we want a Pillow Picture sort, we can not straight use the OpenCV array. To transform an OpenCV array to Pillow, we first have to convert the colour area from BGR to RGB after which we will use the Picture.fromarray perform to create the Pillow Picture object. Subsequent we will create a CTKImage, the place we use the identical picture irrespective of the theme and set the scale in keeping with our body. Lastly we will use the configure methodology to set the picture in our body. On the finish, we additionally reset the textual content variable to take away the “Loading…” textual content, despite the fact that it could theoretically be hidden behind the picture.

To shortly take a look at this, we will set the primary picture of our webcam within the constructor. (We’ll see in a second why this isn’t such a good suggestion)

class App(customtkinter.CTk):
    def __init__(self) -> None:
        ...
        
        cap = cv2.VideoCapture(0)
        _, frame0 = cap.learn()
        self.image_display.set_frame(frame0)

If you happen to run this, you’ll discover that the window takes a bit longer to pop up, however after a brief delay it’s best to see a static picture out of your webcam.

NOTE: If you happen to don’t have a webcam prepared you may as well simply use a neighborhood video file by passing the file path to the cv2.VideoCapture constructor name.

Now this isn’t very thrilling, for the reason that body doesn’t replace but. So let’s see what occurs if we attempt to do that naively.

class App(customtkinter.CTk):
    def __init__(self) -> None:
        ...

        cap = cv2.VideoCapture(0)
        whereas True:
            ret, body = cap.learn()
            if not ret:
                break

            self.image_display.set_frame(body)

Nearly the identical as earlier than, besides now we run the body loop as we did within the earlier chapter with the OpenCV GUI. If you happen to run this, you will note… precisely nothing. The window by no means exhibits up, since we’re creating an infinite loop within the constructor of the app! That is additionally the explanation why this system solely confirmed up after a delay within the earlier instance, the opening of the Webcam stream is a blocking operation, and the occasion loop for the window can not run, so it doesn’t present up but.

So let’s repair this by including a barely higher implementation that permits the gui occasion loop to run whereas we additionally replace the body each now and again. We are able to use the after methodology of tkinter to schedule a perform name whereas yielding the method in the course of the wait time.

        ...

        self.cap = cv2.VideoCapture(0)
        self.after(10, self.update_frame)

    def update_frame(self) -> None:
        """
        Replace the displayed body.
        """
        
        ret, body = self.cap.learn()
        if not ret:
            return
        
        self.image_display.set_frame(body)

        self.after(10, self.update_frame)

So now we nonetheless arrange the webcam stream within the constructor, so we haven’t solved that drawback but. However a minimum of we will see a steady stream of frames in our picture part.

Making use of Filters

Now that the body loop is working. we will re-implement our filters from the start and apply them to our webcam stream. Within the update_frame perform, we will examine the present filter variable and apply the corresponding filter perform.

    def update_frame(self) -> None:
        ...
        
        # Get the chosen filter sort
        filter_id = self.filter_var.get()
        filter_type = filter_types[filter_id]

        if filter_type == "grayscale":
            body = cv2.cvtColor(body, cv2.COLOR_BGR2GRAY)
        elif filter_type == "blur":
            body = cv2.GaussianBlur(body, ksize=(15, 15), sigmaX=0)
        elif filter_type == "threshold":
            grey = cv2.cvtColor(body, cv2.COLOR_BGR2GRAY)
            _, body = cv2.threshold(grey, thresh=127, maxval=255, sort=cv2.THRESH_BINARY)
        elif filter_type == "canny":
            body = cv2.Canny(body, threshold1=100, threshold2=200)
        elif filter_type == "sobel":
            body = cv2.Sobel(body, ddepth=cv2.CV_64F, dx=1, dy=0, ksize=5)
        elif filter_type == "laplacian":
            body = cv2.Laplacian(body, ddepth=cv2.CV_64F)
        elif filter_type == "regular":
            move

        if body.dtype != np.uint8:
            # Scale the body to uint8 if obligatory
            cv2.normalize(body, body, 0, 255, cv2.NORM_MINMAX)
            body = body.astype(np.uint8)
        if len(body.form) == 2:  # Convert grayscale to BGR
            body = cv2.cvtColor(body, cv2.COLOR_GRAY2BGR)
        
        self.image_display.set_frame(body)

        self.after(10, self.update_frame)

And now we’re again to the complete performance of the applying, you possibly can choose any filter on the left aspect and will probably be utilized in real-time to the webcam feed!

Multithreading and Synchronization

Now though the applying runs as is, there are some issues with the present manner we run our body loop. At the moment all the things runs in a single thread, the principle GUI thread. This is the reason to start with, we don’t instantly see the window pop up, our webcam initialization blocks the principle thread. Now think about, if we did some heavier picture processing, perhaps working the pictures by means of neural community, you wouldn’t need your consumer interface to at all times be blocked whereas the community is working inference. It will result in a really unresponsive consumer expertise when clicking the UI parts!

A greater solution to deal with this in our software is to separate the picture processing from the consumer interface. Typically that is virtually at all times a good suggestion to separate your GUI logic from any sort of non-trivial processing. So in our case, we are going to run a separate thread that’s answerable for the picture loop. It can learn the frames from the webcam stream and apply the filters.

NOTE: Python threads aren’t “actual” threads in a way that they don’t have the potential to run on completely different logical cpu cores and therefore won’t actually run in parallel. In Python multithreading the context will swap between the threads, however because of the GIL, the worldwide interpreter lock, a single python course of can solely run one bodily thread. If you would like “actual” parallel processing, you have to to make use of multiprocessing. Since our course of right here will not be CPU certain however truly I/O certain, multithreading suffices.

class App(customtkinter.CTk):
    def __init__(self) -> None:
        ...

        self.webcam_thread = threading.Thread(goal=self.run_webcam_loop, daemon=True)
        self.webcam_thread.begin()

    def run_webcam_loop(self) -> None:
        """
        Run the webcam loop in a separate thread.
        """
        self.cap = cv2.VideoCapture(0)
        if not self.cap.isOpened():
            return

        whereas True:
            ret, body = self.cap.learn()
            if not ret:
                break

            # Filters
            ...

            self.image_display.set_frame(body)

If you happen to run this, you’ll now see that our window opens up instantly and we even see our loading textual content whereas the webcam stream is opening up. Nonetheless, as quickly because the stream begins, the frames begin to flicker. Relying on quite a lot of elements, you may expertise completely different visible artifacts or errors at this stage.

Warning: flashing picture

Now why is that this occurring? The issue is that we’re concurrently attempting to replace the brand new body whereas the interior refresh loop of the consumer interface may be utilizing the data of the array to attract it on the display. They’re each competing for a similar body array.

It’s typically not a good suggestion to straight replace the UI parts from a special thread, in some frameworks this may even be prevented and can increase exceptions. In Tkinter, we will do it, however we are going to get bizarre outcomes. We want some sort of synchronization between our threads. That’s the place the Queue comes into play.

You’re most likely conversant in queues from the grocery retailer or theme parks. The idea of the queue right here could be very comparable: the primary ingredient that goes into the queue additionally leaves first (First In First Out).

On this case, we truly simply desire a queue with a single ingredient, a single slot queue. The queue implementation in Python is thread-safe, that means we will put and get objects from the queue from completely different threads. Good for our use case, the processing thread will put the picture arrays to the queue and the GUI thread will attempt to get a component, however not block if the queue is empty.

class App(customtkinter.CTk):
    def __init__(self) -> None:
        ...

        self.queue = queue.Queue(maxsize=1)

        self.webcam_thread = threading.Thread(goal=self.run_webcam_loop, daemon=True)
        self.webcam_thread.begin()

        self.frame_loop_dt_ms = 16  # ~60 FPS
        self.after(self.frame_loop_dt_ms, self._update_frame)
    
    def _update_frame(self) -> None:
        """
        Replace the body within the picture show widget.
        """
        attempt:
            body = self.queue.get_nowait()
            self.image_display.set_frame(body)
        besides queue.Empty:
            move

        self.after(self.frame_loop_dt_ms, self._update_frame)

    def run_webcam_loop(self) -> None:
        ...

        whereas True:
            ...

            self.queue.put(body)

Discover how we transfer the direct name to the set_frame perform from the webcam loop which runs in its personal thread to the _update_frame perform that’s working on the principle thread, repeatedly scheduled in 16ms intervals.

Right here it’s vital to make use of the get_nowait perform in the principle thread, in any other case if we’d use the get perform, we’d be blocking there. This name does not block, however raises a queue.Empty exception if there’s no ingredient to fetch so we have now to catch this and ignore it. Within the webcam loop, we will use the blocking put perform as a result of it doesn’t matter that we block the run_webcam_loop, there’s nothing else needing to run there.

And now all the things is working as anticipated, no extra flashing frames!

Conclusion

Combining a UI framework like Tkinter with OpenCV permits us to construct trendy trying purposes with an interactive graphical consumer interface. As a result of UI working in the principle thread, we run the picture processing in a separate thread and synchronize the information between the threads utilizing a single-slot queue. You could find a cleaned up model of this demo within the repository under with a extra modular construction. Let me know if you happen to construct one thing fascinating with this strategy. Take care!

Checkout the complete supply code within the GitHub repo:

https://github.com/trflorian/ctk-opencv