Optimization of algorithms for dim light enhancement and object detection in deep learning models for abandoned object recognition

Littering garbage or abandoned object items could have a certain impact on people, and it is not only troublesome to recycle them but also difficult to find the person who threw them away in low light conditions. In recent years, deep learning-based dim light enhancement algorithms and object detection algorithms have gradually emerged as the core of object recognition in image recognition. Developing dim light enhancement algorithms and object detection algorithms is a challenging task due to various problems such as robustness, data limitation, reliability, and computational efficiency. We constructed a dataset of abandoned objects by capturing them from multiple angles and using LabelImg for image annotation. We propose an improved model, ZERO-DCE-CB, to address the issues of poor denoising performance and incomplete feature extraction in addressing the issue of dark light enhancement. For target recognition, the BiFormer with dynamic sparse attention mechanism and step-less convolution (SPD) Conv were combined with YOLOv7-BS to enhance the model’s ability to extract image features. The proposed model has a flexible structure and can change at the training or inference phase to accommodate the amount of available information. The proposed model is evaluated on both the constructed abandoned object dataset and the original dataset. The results showed that ZERO-DCE-CB has a better brightening effect and less noise compared with the original model and other models, with an increase of 0.36 and 0.01 in peak signal-to-noise ratio and structural similarity, respectively. The accuracy of the improved YOLOv7-BS has been improved to 89.9%, and there almost have been no missed detections. Although the focus has been on abandoned objects, the proposed design methodology can be generalized to recognize object patterns in images.