Traditional computer vision techniques involve methods and algorithms that do not rely on deep learning or neural networks. Instead, these approaches are not data-driven and they use classical approaches to process and analyze images. So, in this post, we’ll explore three thresholding techniques!

Thresholding

When the task is to distinguish the background from the foreground, thresholding provides a straightforward solution. We will use this image as an example.

This technique segments an image by assigning one value (typically white) to all pixels above a specified threshold and another value (usually black) to the remaining pixels. Thresholding is a simple yet effective method for separating objects from the background, especially when the background is not complicated.

Code of the histogram:

img = cv2.imread("images/document.PNG")
img = cv2.cvtColor(img, cv2.COLOR_BGR2GRAY)

plt.hist(img.reshape(-1), bins=255, range=(0, 255), zorder=3, color="teal")
plt.xlim([0,255])
plt.ylim(plt.ylim())
plt.vlines(130, *plt.ylim(), label='Threshold', linestyles='dotted', color='red', zorder=3)
plt.grid(alpha=0.3, zorder=1)
plt.title("Histogram of the image")

Binary Thresholding

Binary thresholding is the most simple method of thresholding. Each pixel in the image is compared to a threshold value: if the pixel value is higher than the threshold, it is set to the maximum value (white); otherwise, it is set to the minimum value (black).

In opencv, we almost always use the method cv2.threshold to apply umbralization over an image. In this case, we are setting the threshold to 130 and the maximum value of the image to 255. cv2.THRESH_BINARY is the flag to indicate which kind of thresholding method we want.

threshold = 130
threshold, binary_image = cv2.threshold(image, threshold, 255, cv2.THRESH_BINARY) # The first element of the response is worthless here.

In our case, the image should turn into:

Otsu Thresholding

But… how to find the best threshold? Well, a japanese person named Otsu did a greatjob. Otsu’s thresholding is an automatic method that calculates the optimal threshold value by minimizing the intra-class variance. This technique is particularly useful for images with bimodal histograms.

otsu_threshold, otsued_img = cv2.threshold(img ,0, 255, cv2.THRESH_BINARY | cv2.THRESH_OTSU)

In our case, the image should turn into:

Adaptive Thresholding

What if we have different lightning conditions? Adaptive thresholding is particularly useful in cases where background intensity vary across the image. Instead of a single global threshold, this technique calculates different threshold values for different regions of the image.

Adaptive thresholding typically uses either the mean or Gaussian weighted sum of the neighborhood of each pixel (typical, right?). In OpenCV, we can specify the type of adaptive method with cv2.ADAPTIVE_THRESH_MEAN_C or cv2.ADAPTIVE_THRESH_GAUSSIAN_C. Additionally, a constant C is subtracted from the calculated threshold value to further adjust the results.

adaptive_threshold_img = cv2.adaptiveThreshold(
    img, 255, cv2.ADAPTIVE_THRESH_GAUSSIAN_C, cv2.THRESH_BINARY, 11, 2
)

In this case, adaptative thresholding success thresholding by avoiding background variations.

fig, axs = plt.subplots(3, 1, figsize=(6, 10))
for ax in axs:
    ax.axis('off')

axs[0].imshow(img, cmap='gray'); axs[0].set_title("Original image")

ret, thresh = cv2.threshold(img, 120, 255, cv2.THRESH_BINARY)
axs[1].imshow(thresh, cmap='gray'); axs[1].set_title("Binary thresholding")

thresh2 = cv2.adaptiveThreshold(img, 255, cv2.ADAPTIVE_THRESH_GAUSSIAN_C, 
                                          cv2.THRESH_BINARY, 11, 20) 
axs[2].imshow(thresh2, cmap='gray'); axs[2].set_title("Adaptive thresholding");