Abstract: Aiming at the problem that  the efficiency and performance of image object recognition algorithms on embedded systems were limited due to the small processing speed and memory size of embedded systems, we proposed a high-performance embedded image object recognition algorithm that combined convolutional neural network (CNN) and non negative sparse representation. Firstly, by utilizing CNN to mine embedded image features, parameter sharing and local perception could reduce the model’s parameter count and computational complexity, thereby  improving computational efficiency. Secondly, convolution operation was performed on embedded images by using Roberts cross gradient filter, preliminary feature mining results were obtained by  combining the Sigmoid function operation, and then the non-linear pooling method was used to downsample the results, thereby reducing the dimensionality of feature mining results and completing the image feature mining task. Finally, we used non negative sparse representation to establish a target recognition model, and  solved the coefficient sparse coefficient vector based on 
multiplicative iterative algorithm.  The target area was  determined through kernel function operation and minimum class residue operation. The experimental results show that the F₁ values of each group of image recognition results obtained by the proposed method are stable above 0.98, and the frame rate is high in embedded image target recognition, indicating that the method has the ability to run efficiently on embedded systems while maintaining high-precision recognition performance.

Key words: convolutional neural network, non negative sparse representation, embedded image, multiplicative iterative algorithm