Image-enhancement-EfficientNet-YOLOv3: Visual detection for slag pot transfer in harsh steelmaking plant environments

In this paper, an Image-Enhancement-EfficientNet-YOLOv3 (IEEY) network based on YOLOv3 is proposed to detect the hooking state of main and tail hooks, which is improved from two aspects. Firstly, considering the steelmaking plant's dark, dusty and smoky environment, the image enhancement method combined with steel slag convolutional neural network (SS-CNN) is studied and adopted for image pre-processing in the IEEY network. The SS-CNN network first understands the global content of the image, such as luminance, colour, hue, etc., and then predicts the hyperparameters of six image filters including: exposure, white balance, gamma, tone curve, contrast and sharpening. These filters are used to enhance the image details under dark light and smoke interference. Secondly, the Efficient Channel Attention Net (ECA-Net) is used to improve the Efficient Net target detection network for feature extraction. The EfficientNet-B0 model, which consists of multiple Mobile Inverted Bottleneck Convolutional (MBConv) blocks, is studied. The improved network is used to fuse the feature pyramid structure of YOLOv3. The experimental results show that our IEEY network achieves 99.1% mean average precision and 27.6 frames per second on the test set for six different states. This scheme has been proven effective in actual tests, with the results exhibiting a sufficiently tiny margin of error compared to the test results obtained from the test set.