What is Convolutional Neural Network (CNN)

How It Works

• Convolution: a filter, for example 3 × 3 pixels, slides over the image and outputs one value per position, producing a feature map
• Activation: a non-linear function such as ReLU is applied to each value
• Pooling or strided convolution: feature maps shrink, so deeper layers see a wider area of the image
• Head: the final layers turn the features into an output, such as class scores, boxes or an embedding vector

Worked Example

A 3 × 3 filter on a single-channel image has 9 weights plus a bias, 10 parameters, whatever the image size. A layer with 64 such filters over a 3-channel RGB input has 64 × (3 × 3 × 3 + 1) = 1,792 parameters. A fully connected layer mapping a 224 × 224 RGB image to the same 64 outputs would need 224 × 224 × 3 × 64 + 64, about 9.6 million.

CNNs vs Vision Transformers

• Vision transformers (ViT) split the image into patches and use attention, which relates any patch to any other from the first layer
• CNNs build in locality, so they learn well from smaller datasets and run efficiently on phones and edge devices
• Transformers scale better with very large training sets, which is why most current image-text embedding models, such as CLIP and SigLIP variants, use them
• Hybrid designs, such as ConvNeXt, carry transformer training ideas into a convolutional network

CNNs in Mixpeek

The face identity extractor detects faces with SCRFD, a convolutional face detector, before computing ArcFace embeddings; see face identity. For image and video search, Mixpeek's embedding models are listed in the model catalog, and a custom extractor can run any CNN you have trained.

Common Pitfalls

• Feeding images at a different size or normalization than the network was trained with
• Training a CNN from scratch on a small dataset instead of fine-tuning a pretrained one
• Assuming a classifier trained on product photos works on video frames with motion blur
• Reading accuracy on a balanced test set as accuracy on skewed real traffic