java - android图像识别比对

Question

最新要做个类似博物馆的项目，要用到图片识别比对，具体场景如下：

游客在入馆的时候或者入园之馆（在家里）下载当前博物馆的数据，其中数据主要是展品的缩略图，以及展品的文字，语音介绍。

游客入馆后，看到自己感兴趣的展品可以用打开APP，APP会调用摄像头，对排到的物体做实时采集，并把当前采集到的帧与下载好的缩略图进行比对，找到代表当前展品的缩略图以及文字音视频，进行播放。

我觉得重点在于这个图像识别比对的过程，可能要计算一下当前两个图像的相似度，高于一定数值的，即认为比对成功。但是我不知道有没有什么合适的图像比对算法，大概看了一下，有opencv的，但是貌似要额外在android的机器上装一个40M的apk才行，我个人觉得这有点牛刀小用了，有没有个小巧的就是图像相似度比对的算法呢？

Answer 1

Image comparison technology does not need to be done by yourself. It can be put on the cloud. Baidu, Sina, Tencent, and Alibaba all have good interface implementations.
I think the implementation process should be like this. 1 user downloads the current museum data , 2. The user takes a picture of the physical object in front of him, 3. uploads the picture to the cloud, 4. displays accordingly based on the returned results!

Answer 2

The current mainstream method is to extract features from the pictures in the museum in advance, and then perform offline or online feature extraction on the images collected by the camera, and then compare them, or directly paste a QR code. .

Answer 3

How about javacv?

Answer 4

1. Image recognition is relatively professional. I personally don’t recommend that the poster does it by himself. Of course, if the poster is majoring in digital image processing, that’s another matter.
2. Consider third-party API access. Search on Google and you should find it. Quite a few
3. If the recognition accuracy requirement is not very high, I suggest you read books on digital image processing to find a suitable algorithm

Answer 5

It is recommended to use the ORB algorithm for similarity comparison. If you think the apk takes up too much space, you can look for an ORB algorithm that does not rely on opencv.

Answer 6

Someone from the same lab worked on a museum virtual reality project before, but it felt like nonsense. Image recognition, if it has not been done before, is still very time-consuming and energy-consuming to actually implement it. The application scenarios you mentioned were probably thought up by decision-makers. They look cool, but they are useless. It might be more practical if you can directly hang a QR code as mentioned above. Many museums do this. But if the poster only has the option of image recognition, it is recommended to find relevant APIs or outsource it to other institutions. If you have plenty of time, insufficient funds, and cannot find a suitable API, you can only use digital images yourself. Processing, matrix operations, various transformations...just treat it as a math game...