Why CNNs dominate image classification in Computer Vision - Why Metrics Matter

Metrics & Evaluation - Why CNNs dominate image classification

Which metric matters for this concept and WHY

For image classification using CNNs, accuracy is the most common metric because it shows how many images are correctly labeled out of all images. However, when classes are imbalanced, precision and recall become important to understand if the model is good at finding specific classes without too many mistakes.

Confusion matrix example

      | Predicted Cat | Predicted Dog |
      |--------------|---------------|
      | True Cat: 90 | False Dog: 10 |
      | False Cat: 5 | True Dog: 95  |

      Total samples = 200

      Precision (Cat) = TP / (TP + FP) = 90 / (90 + 5) = 0.947
      Recall (Cat) = TP / (TP + FN) = 90 / (90 + 10) = 0.9
      Accuracy = (TP + TN) / Total = (90 + 95) / 200 = 0.925
    

Precision vs Recall tradeoff with examples

In image classification, sometimes you want high precision to avoid false alarms. For example, if a model detects rare animals, you want to be sure when it says "this is a rare animal" it is correct (high precision).

Other times, you want high recall to catch all instances. For example, in medical image classification, you want to find all tumors even if some false alarms happen (high recall).

CNNs help balance this tradeoff by learning detailed features that improve both precision and recall compared to older methods.

What "good" vs "bad" metric values look like for this use case

Good: Accuracy above 90%, precision and recall both above 85% means the CNN is correctly classifying most images and not missing many.

Bad: Accuracy around 50-60% or precision very low (below 50%) means the model guesses poorly or confuses classes a lot.

Common pitfalls in metrics for CNN image classification

• Accuracy paradox: High accuracy can be misleading if one class dominates the dataset.
• Data leakage: If test images are too similar to training images, metrics look better than reality.
• Overfitting: Very high training accuracy but low test accuracy means the CNN memorized training images but can't generalize.

Self-check question

Your CNN model has 98% accuracy but only 12% recall on a rare class like cancer in images. Is it good for production? Why or why not?

Answer: No, it is not good. The model misses 88% of cancer cases (low recall), which is dangerous. High accuracy is misleading because cancer cases are rare. You need to improve recall to catch more cancer cases.

Key Result

Accuracy is key for overall performance, but precision and recall reveal CNN strengths in correctly identifying image classes and handling imbalanced data.

Practice

1. Why are Convolutional Neural Networks (CNNs) especially good for image classification?

easy

A. Because they only work with black and white images

B. Because they use random guessing to classify images

C. Because they ignore image details and focus on text

D. Because they scan small parts of images to find important patterns

Solution

1. Step 1: Understand CNN scanning method

   CNNs look at small parts of an image called patches to detect patterns like edges or shapes.

2. Step 2: Connect scanning to image classification

   By scanning patches, CNNs learn important features that help tell one image from another.

3. Final Answer:

   Because they scan small parts of images to find important patterns -> Option D

4. Quick Check:

   CNN scanning = small parts pattern detection [OK]

Hint: Remember CNNs focus on small image parts to find patterns [OK]

Common Mistakes:

• Thinking CNNs guess randomly
• Believing CNNs ignore image details
• Assuming CNNs only work on black and white images

2. Which of the following is the correct way to describe the pooling operation in CNNs?

easy

A. Pooling increases the image size to add more details

B. Pooling shrinks the image while keeping important information

C. Pooling removes all colors from the image

D. Pooling randomly changes pixel values

Solution

1. Step 1: Define pooling in CNNs

   Pooling reduces the size of the image or feature map but keeps the key features intact.

2. Step 2: Identify correct description

   Pooling does not increase size or remove colors; it shrinks the image while preserving important info.

3. Final Answer:

   Pooling shrinks the image while keeping important information -> Option B

4. Quick Check:

   Pooling = shrink + keep key info [OK]

Hint: Pooling shrinks images but keeps what matters [OK]

Common Mistakes:

• Thinking pooling makes images bigger
• Believing pooling removes colors
• Assuming pooling changes pixels randomly

3. Given this simple CNN layer code snippet in Python using PyTorch:

import torch
import torch.nn as nn
conv = nn.Conv2d(in_channels=3, out_channels=1, kernel_size=3)
input_tensor = torch.randn(1, 3, 5, 5)
output = conv(input_tensor)
print(output.shape)

What will be the shape of the output tensor?

medium

A. torch.Size([1, 1, 3, 3])

B. torch.Size([1, 3, 3, 3])

C. torch.Size([1, 1, 5, 5])

D. torch.Size([3, 1, 3, 3])

Solution

1. Step 1: Understand Conv2d output size formula

   Output size = (Input size - Kernel size + 1) for default stride and padding. Here, input is 5x5, kernel is 3x3, so output is 3x3.

2. Step 2: Check channels and batch size

   Batch size is 1, output channels is 1, so output shape is (1, 1, 3, 3).

3. Final Answer:

   torch.Size([1, 1, 3, 3]) -> Option A

4. Quick Check:

   Output shape = (1, 1, 3, 3) [OK]

Hint: Output size = input - kernel + 1 with default stride [OK]

Common Mistakes:

• Confusing input and output channels
• Forgetting batch size dimension
• Assuming output size equals input size

4. Identify the error in this CNN pooling layer code snippet:

import torch
import torch.nn as nn
pool = nn.MaxPool2d(kernel_size=2, stride=3)
input_tensor = torch.randn(1, 1, 6, 6)
output = pool(input_tensor)
print(output.shape)

What is the problem with this code?

medium

A. Input tensor shape is invalid for pooling

B. Kernel size must be equal to stride in MaxPool2d

C. Stride is larger than kernel size, causing unexpected output size

D. MaxPool2d does not accept stride as a parameter

Solution

1. Step 1: Check pooling parameters

   Stride can be different from kernel size, but stride larger than kernel size can cause skipping regions and smaller output.

2. Step 2: Understand effect on output size

   Stride 3 with kernel 2 on 6x6 input reduces output size more than expected, which may cause loss of important info.

3. Final Answer:

   Stride is larger than kernel size, causing unexpected output size -> Option C

4. Quick Check:

   Stride > kernel size affects output size [OK]

Hint: Stride bigger than kernel skips image parts, watch output size [OK]

Common Mistakes:

• Thinking kernel size must equal stride
• Believing input shape is invalid
• Assuming MaxPool2d can't take stride

5. You want to build a CNN that classifies images of cats and dogs. Which combination best explains why CNNs dominate this task compared to a simple fully connected network?

hard

A. CNNs scan local image parts and use pooling to reduce size, capturing patterns efficiently

B. Fully connected networks scan images in small parts and pool features

C. CNNs ignore image patterns and rely on random weights

D. Fully connected networks use convolution layers to find edges

Solution

1. Step 1: Compare CNN and fully connected networks

   CNNs scan small parts of images (local receptive fields) and use pooling to keep important info while reducing size.

2. Step 2: Understand why CNNs are better for images

   Fully connected networks treat all pixels equally without spatial structure, making them less efficient for images.

3. Final Answer:

   CNNs scan local image parts and use pooling to reduce size, capturing patterns efficiently -> Option A

4. Quick Check:

   CNN local scan + pooling > fully connected for images [OK]

Hint: CNNs scan parts + pool; fully connected treats all pixels equally [OK]

Common Mistakes:

• Confusing fully connected with convolution layers
• Thinking CNNs ignore image patterns
• Believing fully connected networks use pooling