Automatic generation of neural networks for image processing

This paper presents a technique for automatic generation of image processing architectures based on artificial neural networks (NN) for real time vision applications in order to reduce the hardware design effort. The generated datapath can be reused with different functions. A high throughput is obtained with one output pixel being produced at each clock cycle for each input pixel, allowing VGA stream processing. NN used is MLP, trained by back-propagation. Function training is executed in a C++ software. Then VHDL code of the image processing IP core is automatically generated. Image processing systems using the generated IP cores were evaluated in FPGA, showing both good performance and suitability of the method