OPTIMIZING OBJECT DETECTION PERFORMANCE THROUGH IMAGE ENHANCEMENT FOR BEE IMAGES

Neural networks and deep learning, being among the technologies experiencing daily growing interest, have found practical application, particularly in bee monitoring. The use of neural networks for identifying bees in images can provide non-invasive monitoring of the condition of individual bees and, consequently, reduce stress on the colony. However, the accuracy of object detection algorithms can be significantly affected by variations in monitoring conditions, such as weather conditions, lighting, or the use of different cameras. Thus, image enhancement can be an important pre-processing step for optimizing object detection algorithms.This study utilizes various image enhancement and color normalization methods to improve the performance of the SSD MobileNet model in detecting the full bodies of bees in images. The data source consists of photos taken with a smartphone, which were subsequently divided into sections to fit within the model's acceptable identification limits. The study examines the effectiveness of three image enhancement methods: Contrast Limited Adaptive Histogram Equalization (CLAHE), a regular histogram equalization algorithm, and Protobuf pipeline functions. Each approach is evaluated on different image sets, including photos of normal beehive sections taken with a smartphone, photos with randomly altered brightness and contrast, and photos with slight blurring. The findings show that image normalization alone does not improve object detection performance on standard bee photos, whereas the CLAHE algorithm proved most effective in other cases, maintaining good results when image brightness and contrast were not optimized. Additionally, it was found that applying blur did not result in an overall increase in the efficiency of identifying bees in photos.