Object detection is a fundamental task in computer vision. To improve the detection accuracy of a detection model without increasing the model weights, this paper modifies the YOLOX model by first replacing some of the traditional convolution operations in the backbone network to reduce the parameter cost of generating feature maps. We design a local feature extraction module to chunk the feature maps to obtain local image features and a global feature extraction module to calculate the correlation between feature points to enrich the feature, and add learnable weights to the feature layers involved in the final prediction to assist the model in detection. Moreover, the idea of feature map reuse is proposed to retain more information from the high-dimensional feature maps. In comparison experiments on the dataset PASCAL VOC 2007 + 2012, the accuracy of the improved algorithm increased by 1.2% over the original algorithm and 2.2% over the popular YOLOv5.

To address the insufficient recognition stability of answer sheets in applications with scanning skew, low resolution, handwriting differences, corrections, and cross-template applications, this paper proposes an intelligent recognition and error localization method based on multimodal large model and OCR fusion. The method first denoises, binarizes, corrects skew, and normalizes perspective on the input image. Then, it detects the page layout to obtain the question number area, multiple-choice question area, fill-in-the-blank question area, subjective question area, and correction area. Different areas are input into the OCR module to obtain candidate text or filling status. Then, the local image blocks, OCR candidates, question type hints, and question number context are fed into the multimodal large model for consistency verification. To avoid misjudgment by a single model, this paper constructs a joint decision function that integrates OCR confidence, multimodal consistency, and page layout matching. The final answer is parsed into a structured result containing the question number, answer, coordinates, confidence, and question type. Furthermore, the system performs difference judgment on multiple-choice, fill-in-the-blank, and subjective questions based on the standard answers, and traces the error results back to the original image coordinates, outputting the error area, error type, and location reliability. Experiments were conducted to construct standard scan, skew scan, fuzzy scan, error correction interference, and cross-template answer sheet samples, comparing them with traditional OCR, template matching + OCR, layout detection + OCR, and a standalone multimodal model. Results show that the proposed method achieves an answer recognition accuracy of 96.8% and an error location F1 score of 95.1%, maintaining a recognition accuracy of 91.8% even under heavy interference conditions, indicating that this method is suitable for intelligent marking and structured examination data collection scenarios.