用于图像到特征向量的 CNN

Question

I have a rather simple question, but I cannot wrap my head around this. You need to correct me in the case that I have understood something wrong,,. Typically, CNNs are used for image classification, where the FC layers are used to output the probability per given class.

I want actually something else: Can I train a CNN to output a feature vector? One example would be: Input==Image; Output==Color histogram with 3 channels (RGB) and every channel value is the number of red, green, blue pixels. (I actually want to learn other image feature descriptors for special images, but that should then work analogous).

Pointers, Links and Tutorials (esp. code that I can modify for my purposes) would be greatly appreciated.

Thanks so much in advance!!

Answer 1

the output of a CNN is a feature vector,it means the input is an image and you'd get a feature vector of that image as an output. That feature vector has been computed by using a mask that apply on that image. Those feature vector is been used for classification. it's up to you that use the output for classification or anything else. this link might help you more. http://www.wildml.com/2015/11/understanding-convolutional-neural-networks-for-nlp/