Pedestrian attribute recognition has been a challenging task due to the small size and low resolution of the surveillance images used, and the low-resolution of the images often leads to a dataset with problems such as indistinct main part of recognized pedestrians and serious background noise interference. Most of the previous methods use unprocessed raw images as inputs, so the attribute recognition results are often unsatisfactory. Moreover, the mainstream datasets for attribute recognition usually suffer from imbalanced positive and negative samples. For example, many pedestrians manifest seasonal or customary deviations in the distribution of clothing attributes. Therefore, this study proposes a new deep learning network, namely, the image enhancement and sample balance optimization (IEBO) model. In this model, methods of color enhancement and noise suppression in the main part of the extracted pedestrians are applied to highlight the core features of the pedestrians and eliminate useless background information to prevent it from interfering with attribute recognition. In addition, the model is optimized through weight adjustment to deal with sample imbalanced attributes and thereby improve the recognition of imbalanced attributes. The experiments show that the new pedestrian attribute recognition model achieves better performance on the Market-1501-attribute dataset.