Knife Detection System Using CLIP-Based Vision-Language Model

Knife detection in restricted areas is essential to ensure security and preventing harm. In institutions such as schools, airports, prisons, or public transport, restricted areas can easily turn into a threat if someone carries a knife. Knifes are detected to prevent possible threats, eliminate injuries, and provide a safe environment for all. AI based Knife detection ensures safety, preventing harm, saves lives, promotes safe public spaces. In this work a machine learning algorithm known as ViT B/32 a Contrastive Language–Image Pretraining (CLIP) based model is presented for efficient detection of knifes in small spaces utilizing the Vision Language Models (VLM). CLIP is a zero-shot detection model, it can classify and interpret data without training, testing, and validation. CLIP leverages pre- trained knowledge from large-scale textual and visual data to make accurate predictions directly. This eliminates the need for the typical data preparation and segmentation process, making it highly efficient for tasks across diverse domains. It examines an image of a knife using a vision encoder to identify its shape, dimensions, and texture as patterns (feature embeddings). The proposed model has been tested with three different knife image datasets. The performance metrics considered for the simulation are precision, recall, F1 score, and accuracy. The dataset1 consists of 400 knife images, dataset2 has 4,075 knife images and dataset3 comprises 3,985 knife images. The proposed algorithm Vit - B/32 outperforms exceptional results with an average precision of 97.67%, accuracy of 89.67%, recall of 84.33%, and F1-score of 89.33% for all the three datasets. A comparison with existing algorithms highlights the effectiveness of the proposed algorithm.