Behavior Recognition Algorithm for Electric Power Workers in Hazy Environments

This paper addresses the issue of reduced recognition accuracy in the field of electric power worker behavior recognition due to the impact of hazy environments, which cause the images to become blurry and features to become less distinct. To tackle this problem, an algorithm based on Retinex algorithm and wavelet transform for haze removal, followed by the use of an improved YOLOv8 model for worker behavior recognition, is proposed to enhance the model's recognition capabilities in hazy environments. Initially, the input RGB image is converted to a YIQ image, and the Y component is extracted. Then, wavelet decomposition is applied to the Y component to separate the low-frequency component containing haze information and the high-frequency component containing image details. The Retinex algorithm is used for haze removal on the low-frequency component, while the high-frequency part is processed with a wavelet threshold function using a hard threshold function to suppress useless parts and enhance useful parts, preserving more features of the image. Finally, the processed low and high-frequency components are reconstructed through wavelet reconstruction to obtain the output image. This processed image is then used as input for the improved YOLOv8 model to recognize the specific behavior characteristics of electric power workers in the image. The specific improvements to the YOLOv8 model are as follows: the standard convolution in the original YOLOv8 bottleneck is replaced with deformable convolution, and the attention mechanism CBAM is added to construct a new bottleneck_den module, which replaces the bottleneck module in C2f of YOLOv8 and is renamed as C2f_DeN. Secondly, all standard convolutions in the backbone of YOLOv8 are replaced with deformable convolutions to enhance the model's ability to extract image features. Experiments show that the improved algorithm model has an mAP of 85.2%, which is an increase of 6.3 percentage points and 4.5 percentage points compared to two other models that have also been processed.