卷积神经网络在二值图像上运行更快吗？手写单词识别图像尺寸均衡咨询

1. Computational Speed: Binary vs Grayscale/Color Images #

Let’s break this down based on the type of DCNN you’re using:

Standard Floating-Point DCNNs #

• If you’re working with a typical DCNN (using floating-point weights and activations), binary images (1 channel) will perform nearly identically to grayscale images (also 1 channel) in terms of inference/training speed—since both use single-channel inputs. The real gap comes when comparing to color images: color inputs (3 channels) require 3x more computation in the first convolutional layer (each kernel has 3 input channels instead of 1), so binary/grayscale will be noticeably faster here.
• Binary images also have smaller file sizes, which speeds up data loading and preprocessing—this can be a huge time-saver when working with large word spotting datasets.

Binary-Specialized DCNNs #

If you use purpose-built binary networks like BinaryNet, XNOR-Net, or BinaryConnect, the speed difference becomes dramatic. These networks use binary weights and activations (only 0s and 1s), replacing expensive floating-point operations with lightweight bitwise operations. This can lead to 10-50x faster inference compared to standard DCNNs, while still retaining enough accuracy for word spotting (since binary images already capture the core shape of handwritten words).

2. Balancing Image Sizes After Normalization #

Handwritten words have wildly varying aspect ratios, so forcing them into a fixed shape without preserving proportions will distort critical letter features (like width, height, and spacing). Here are the most practical approaches:

• Resize + Aspect Ratio Preservation + Padding
  This is the industry standard for fixed-input DCNNs:

  1. Pick a target dimension (e.g., 64px height, 256px max width) based on your dataset’s typical word sizes.
  2. Resize the image so its height matches the target height (or width matches the max width if the word is extra wide), keeping the original aspect ratio intact.
  3. Pad the empty space (left/right or top/bottom) with a neutral value (e.g., 0 for binary images, since they’re black-and-white). This ensures all inputs are the same fixed size for your DCNN, while preserving the word’s original shape.

• Adaptive Pooling in the Network
  If you want to avoid padding entirely, modify your DCNN to handle variable-sized inputs:

  • Add a global adaptive average/max pooling layer right before the fully connected layers. This layer automatically resizes the feature map to a fixed size (e.g., 64x64) no matter what the input image dimensions are.
  • This works great for fully convolutional networks (FCNs) and eliminates padding artifacts, though you’ll need to ensure your framework (PyTorch/TensorFlow, etc.) supports variable input sizes (most modern ones do).

• Fixed Height, Variable Width
  For architectures using recurrent layers (like LSTMs) common in sequence-based word spotting, you can normalize all images to a fixed height (e.g., 32px) and leave the width variable. The network processes each column of the image sequentially, naturally handling short and long words without distortion. This is ideal for preserving full word context.

Quick Tip #

Never stretch or squish images to fit a fixed size—this warps letter shapes and will tank your model’s accuracy. Always prioritize aspect ratio preservation unless you’re using a network explicitly designed to handle distorted inputs (which is rare for word spotting).

内容的提问来源于stack exchange，提问作者innuendo