This paper analyzes public interest litigation and its salient features, and organizes the audit rules for the electronic transformation of litigation evidence. Aiming at the phenomenon of varying text length in litigation evidence, a joint CTC-Attention decoding model (HCADecoder) based on bigram hybrid labeling is proposed. Based on the existing research on computer vision technology for target number prediction, the stacked object occlusion problem existing in special scenes is proposed, and an algorithm for predicting the number of stacked objects combining planar density map and depth map is proposed. Combined with the public interest litigation evidence document corpus dataset, we analyze the recognition of basic elements of litigation evidence by text label recognition algorithm, and select the commonly used precision rate P, recall rate R and F1 value to evaluate the recognition results of basic elements. Subdivide the text length of litigation evidence and analyze the recognition accuracy of each algorithm on different text lengths. Bring the text label recognition algorithms into real cases to analyze the element extraction. For this paper, we propose monocular image target counting algorithm, which is brought into different scenarios for performance testing. This paper proposes text label recognition algorithm with evidence image target counting algorithm for litigation evidence text image recognition with mean value at 80%.