Learning orientation detection system and its application to RGB images
Tianqi Chen, Zhiyu Qiu, Yuxiao Hua, Yuki Todo, Zheng Tang · 2024
Our previous research showed that emulating dendritic neuron structure effectively addresses orientation detection challenges in learning tasks, reducing both learning time and costs compared to alternative neural network approaches. Our simulation model incorporates an On-Off Response mechanism in bipolar cells (BCs) and horizontal cells (HCs) for processing grayscale input images. To overcome limitations with single-channel input, we introduce color-selective cells. By integrating these cells, we enhance and select outputs of local orientation detection dendritic neurons, generating specific feature maps. These feature maps are generated using a biologically-inspired model that mimics the mechanism of color-selective cells. Additionally, global neurons are used to capture overall image features by aggregating outputs from local dendritic neurons. Our system utilizes backpropagation to update parameters of local orientation detection dendritic neurons. Furthermore, we integrate a learnable orientation detection neural network after the dendritic neuron stage.