Abstract: Single pixel imaging has been widely used in many fields due to its advantages of low cost, flexibility, and strong anti-interference ability. However, commonly used orthogonal basis functions such as Fourier, wavelet, Hadamard, etc. cannot fully capture the features of complex images and the loss of spectral coefficients during the transformation process, resulting in poor reconstruction quality. In order to improve the imaging quality of single pixel imaging based on Hadamard transform, this paper proposes a Hadamard single pixel imaging reconstruction method based on an improved U-net algorithm. The algorithm combines the spectral characteristics of Hadamard single pixel imaging to construct a loss function, and introduces adaptive pooling and attention mechanisms to enhance input features. Through channel attention and spatial attention mechanisms, the model's response to important features is dynamically enhanced. The experimental results show that at a low sampling rate of 6.25%, compared with traditional Hadamard single pixel imaging of butterfly samples, the peak signal-to-noise ratio (PSNR) is improved by about 12%, and the structural similarity (SSIM) is improved by about 30%. Compared with the Hadamard single pixel imaging method based on SRCNN, the PSNR is improved by about 4% and the structural similarity (SSIM) is improved by about 3%.It can achieve good reconstruction tasks for single pixel imaging at ultra-low sampling rates.