CRNNs

Convolutional Recurrent Neural Networks (CRNNs) are neural network architectures that combine the feature extraction capabilities of convolutional neural networks with the sequence modeling strengths of recurrent neural networks. They are designed for data that have both spatial structure and temporal or sequential content, such as lines of text in images or sequences of video frames.

In a typical CRNN, a stack of convolutional layers processes the input to produce a sequence of feature vectors along a chosen axis, usually the horizontal width of a text line. One or more recurrent layers, such as LSTM or GRU units, then process this sequence to capture temporal dependencies, often in both directions. The network ends with a dense layer and a softmax over output symbols. Bidirectional recurrent layers are common to incorporate contextual information from both past and future positions. Training frequently uses Connectionist Temporal Classification (CTC) loss, which aligns unsegmented label sequences with the predicted sequence without requiring exact character boundaries. Decoding can employ greedy methods or beam search.

Applications are dominated by optical character recognition and scene text recognition, including handwriting recognition and text extraction from images, with extensions to video activity recognition where frame-based features are sequenced over time. Variants include attention-based CRNNs, as well as architectures that replace RNNs with temporal convolutions or Transformers for efficiency or performance.

Strengths include end-to-end trainability, handling of variable-length outputs, and no need for pre-segmented labels. Limitations involve computational demand and reliance on large, labeled datasets to achieve robust performance.