Image Classification

Robustness 鲁棒性

Edge images

Edge Detection
有关边缘检测算法

Image Classification:

input: image
output: Assign image to one of a fixed set of categories.
Problem: Semantic Gap(语义差距)
Object Detection:

One way to perform is via image classification of different sliding windows im the image. In a nutshell, it means classify different sub-regions of the image so we could look at s sub region over here and then classify it as background, horse, person, car or a truck…

1
2
3
def classify_image(image):
	# Some magic here?
	return class_label

Machine Learning:Data-Driven Approach

  1. Collect a dataset of images and labels
  2. Use Machine Learning to train a classifier
  3. Evaluate the classifier on new images
    1
    2
    3
    4
    def train(images, labels):
    	# Machine learning
    	# Memorize all data and labels
    	return model
1
2
3
4
def predict(model, test_images):
	# Use model to predict labels
	# Predict the label of the most similar training image
	return test_labels

Distance Metric to compare images

曼哈顿距离:

L1(Manhattan) distance: $d_1(I_1,I_2) = \sum_p|I_1^p-I_2^p|$


Nearest Neighbor Decision Boundaries

Decision Boundarie is the boundary between two classification regions.

Decision Boundarie can be nosiy; affected by outliers.

K-Nearest Neighbors

Using more neighbors helps smooth out rough decision boundaries.

Using more neighbors helps reduce the effect of outliers.

When K > 1 there can be ties between classes. Need to break somehow!


白色区域均有三个最近的邻居。

欧几里得距离

L2(Euclidean) distance = $d_2(I_1, I_2) = \sqrt{\sum_{p}(I_1^p - I_2^p) ^ 2}$

Setting Hyperparameters

Among the three ideas above, The best idea is Idea #3
image.png

当然，我们也可以将数据集划分为更多，以此更好估计我们的泛化性能。
也就是交叉验证
image.png
随着训练的数量越来越多




We can see the nearest neighbor classifier performs well.