Convolutional neural network (CNN)has been widely used to improve traditional neural networks.However,as the performance of the CNN improves the size of its model increases,which requires larger storage and more computing po wer,so the processors based on the von Neumann architecture is difficult to achieve satisfa ctory processing performance.In order to improve the performance of the system,near memory comp uting (NMC) is a promising research direction at present.In this pap er,a reconfigurable array processor-based NMC structure is used to implement handwritten digit recognition,and convolution operations are realized in parallel;at the s ame time,the shared cache array structure is used to reduce off-chip storage access.Experim ental results show that with the operating frequency of 110 MHz,the calculation speed of a single 5 ×5 convolution operation is increased by 75.00%,and a handwritten digit can be identified with in 9 960 μs.