Images are undoubtedly the most important resource medium on the internet, and search engines are undoubtedly the most important channel for acquiring resources online. Naturally, when images meet search engines, they quickly spark productization.Currently, almost all mainstream internet search engines provide image search functionality (prior to this, there were specialized vertical search engines like Tineye and Picitup). The most common form is through users inputting keywords to retrieve related images. However, few people notice another type of search: reverse image search. On the Google Image Search interface, clicking on the camera icon to the right of the search bar (on Baidu, it's a picture icon) allows users to upload an image or reference an image link. The search engine then searches based on the content of the image. But why is this seemingly cool feature not commonly used, or even rarely known?Unsatisfactory search resultsImage search engines mostly use image fingerprinting to represent images, which limits the target images retrieved to be basically consistent with the original image or at least largely similar. Scaling transformations of the target image are allowed on this basis. It can be imagined that users rarely need to obtain another image identical to the one they uploaded.Content retrieved is less than satisfactoryLet's try to guess what kind of results users want when they upload an image? There could be scenarios like this:Student A: Uploads a celebrity's avatar, urgently wanting to know who she is and everything related to her. In fact, to a large extent, if we know who she is, we can retrieve everything about her on the internet using her name or nickname.Google's approachAnalyze web pages containing the target image extensively to identify the keyword most relevant to the image, and use this keyword to enhance the search. In this scenario, the keyword happens to be the celebrity's name.Baidu's approachFurther, when there is obvious facial content in the image, facial recognition will be initiated, returning images containing similar facial content.Sogou's approachBuilding on Google's method, Sogou introduced themed image collections, which include potentially related but not identical images.In this scenario, Student A's problem has been satisfactorily resolved. But let's also take a look at Student B.Student B: Uploads an image of a sausage tree, trying to find more background information or images about this plant.Google's approachThis time, Google cannot obtain its keyword, making it even harder to infer the user's search intent. Fortunately, it returns some images that are similar in content, especially in color, but most of them are not the sausage tree images Student B wants.Baidu's approach (same as Sogou)Baidu cannot obtain the keyword, nor can it guess the user's search intent, so it essentially stops working.User needs are complex, and with current technology, it's almost impossible to accurately infer user needs solely relying on image recognition.What can be doneIf that's the case, what can users do with reverse image search? Sogou provides a brief example:Sogou's explanation page also hints that many users are still unaware of reverse image search, and even less clear about how to utilize it. Meanwhile, internet product managers have no clear product application positioning, only briefly mentioning some specific application scenarios, which are more akin to the research scope of vertical search engines.Peeling away the surface-level functional packaging, it can be analyzed that through images as a medium, the following types of internet resources are easiest to obtain:The awkward truth is that whether these resources are useful to users still requires the users themselves to judge.Image recognition is not the ultimate goalJust as search engines cannot fully parse human natural language, image recognition technology cannot fully understand the semantic content of images, let alone the semantics users assign to images. Perhaps with the current pattern recognition technology, achieving vertical search through reverse image search is indeed quite difficult. Instead, clearly listing the most easily obtainable resources and letting users screen and interact may be more feasible and practical.However, as long as there are still people holding a photo asking which singer it is, reverse image search deserves investment in research. As geographic information and multimedia media technologies continue to develop, search engines will be able to acquire more user background information, associate more media category resources, strengthen the overall resource network connections, and achieve more ideal feedback results.Additionally, perhaps on mobile devices, visual search might have significant potential. Because the shift from Web to App-based search better aligns with people’s anytime, anywhere search habits. Due to changes in context, user needs also change, and it can also be integrated with social networks and other Apps on mobile devices. As previously mentioned by kaler, an observer at GeekPark, if you see a tie in a store that you like, but you don't like the color, and you hope for a cheaper price on Taobao, you can just take a picture of the tie and jump directly to Taobao. You'll immediately get the exact price and information about other similar ties, giving you more choices and allowing you to quickly select products that truly meet your needs.