The rapid growth of regional language news videos faces major challenges for efficient indexing and retrieval, particularly for regional languages such as Gujarati, which is spoken by 55 to 62 million people worldwide. Processing every frame in videos is time-consuming and redundant; therefore, accurate shot boundary detection followed by effective key-frame selection is important for text-based video retrieval. This research presents an adaptive video segmentation framework for Gujarati news videos that integrates shot boundary detection with key frames selected using statistical methods. Initially, consecutive video frames are processed category-wise, and six frame-level features are extracted from the videos, such as PDM, CDM, HBA, Gabor response, ECR, and EED. These features are then normalized and fused using a weighted approach. Mean and standard deviation based adaptive thresholds are then applied to detect both abrupt and gradual shot transitions. For the detected shots, a key frame has been extracted from shots containing Gujarati text based on the Ground truth table. This approach uses entropy and edge density to capture detailed information and structural variation. Adaptive thresholds derived from their mean values are used to preserve descriptive frames with high textual significance. Experimental evaluation on TV9 Gujarati news video datasets shows that the proposed framework effectively reduces redundant frames while preserving semantically significant content key frames, which are suitable for efficient Gujarati text-based video retrieval. Total no. of shots reduced is 2158 from 80675, and total key frames are 901, with the lowest accuracy in the weather category, with 81.0714 % and the highest recorded as 88.00 % for the Cricket category.