Commodity retrieval is fundamental to the intelligent development of the e-commerce industry. This study is concerned with a fashion recognition system which bases on ZYNQ and CNN model, trains the custom network with TensorFlow and processes the weighs by using fixed-point calculations. This system applies ARM + FPGA software and hardware coordination method of the ZYNQ device to construct its framework, ARM to preprocess image by OpenCV and the CNN IP of FPGA to synchronously recognize the image. A weight-reloadable structure is implemented between ARM and FPGA, so online upgrade could be realized without modifying the FPGA hardware. The system uses fashion-minist datasets as the sample in the network training and improves the parallelism of convolution operations by increasing the acceleration engines of CNN IP. According to the experiment, the system realizes accurate and real-time identification and display for actual pictures of the e-commerce platform and its accuracy reaches 92.39%. The image processing speed can reach 1.261 ms per frame and the power consumption is only 0.53 W at 100 MHz.