You can find all the datasets, experiments, and results in the Github repository. Finally, through experiments, we concluded the best methods according to the requirements: speed and accuracy. Also, we determined the sensitivity of image transformations on the performance of those algorithms. In conclusion, we walked through the concepts and Python code of Euclidean distance, SSIM, image hashing, cosine similarity, and the similarity of features. Visit this article to learn more about when Euclidean distance is better than Cosine similarity and vice versa. Moreover, Image hashing is a great candidate if speed is a huge factor. The results (shown in Table 1) clearly show that cosine similarity dominates while computing the similarity of features using CNNs is considered over-engineering for finding duplicates/near-duplicates of an image since its running time is around 250x slower than cosine similarity while keeping up close accuracy. Calculate the accuracy (for each reference image, if the 3-duplicates/near-duplicates are detected, the accuracy is 100%).

Calculate the average time in seconds the method took to compare a pair of images.

Get the 3-most similars for a reference image.

Read the images of the Fruits360 dataset.

This test proposes the best method in terms of speed and accuracy for the image duplicate finder system. Also, you can find all the tests in the Github repository. Note: I used a 2019 MacBook Pro CPU for all the tests.

Experiment 2: Resilience to Image Transformation.

I tested each method using both datasets in the following manner: Image 13 - A Sample of the SFBench dataset (image by author) Experiments

Image duplicate finder