Search can be said to be a very good functional design for users. The more search algorithm data we can provide, the more accurate the results will be. Today, let's take a look at several implementation methods of search algorithm.

1. Understand the data and use stratified sampling.

Data is the core of the whole process strategy. Image quality may bring complexity and need to be handled carefully. In visual search, it is very important to know the pictures in the directory and the pictures uploaded by users.

2. Data enhancement is the key, especially image rotation.

When the training data can't contain all possible changes in the real scene, data expansion is a key step in training neural networks. When a user shoots an object with a mobile phone, the image is likely to be cropped, rotated, blurred or not scaled to an appropriate size.

3. Use supervision information as much as possible when extracting semantic signatures.

As mentioned above, it is very important to use supervision information as much as possible. This helps to train classifiers to pay attention to information content and ignore other non-information areas.

4. Entropy analysis of signature

This step is usually ignored in the system design of large-scale information retrieval systems. It is very important to evaluate whether a given signature capacity contains enough valid information.

5. When the label is a coarse-grained label, within-class variance is very important.

We use coarse-grained leaf category labels instead of product ID to train neural networks. Part of the reason is that although the leaf category is coarse-grained, it is easier to obtain.

6. Use the exclusion method to improve the search speed and accuracy.

Exclusion method specially designed for high speed and high precision has powerful ability. For example, if the input image contains sneakers, there is no need to search the inventory of skirts, tables and computers.