Although the traditional convolutional neural network has good application accuracy, its main defect is low efficiency. In order to solve this problem, the weak supervised learning algorithm is proposed. The existing weak supervised learning algorithm has less labeled training samples and ideal efficiency, but there is still a lack of high misclassification rate. In order to meet the requirements of high efficiency and high precision at the same time, this study combines weak supervision algorithm and convolutional neural network, a weak supervised convolutional neural network feature learning algorithm based on class space constraints is proposed. Firstly, the network model of the feature learning algorithm of weakly supervised convolutional neural network was established; secondly, by constraining the space, the labeled samples and unlabeled samples were connected to realize the feature space clustering; finally, the training sample data was used to realize the design of weak supervised convolutional neural network feature learning algorithm based on class space constraint. The results show that the proposed method has a misclassification rate of 5% and a classification time is no more than 0.4ms, which can better carry out feature learning.