Lightweight underwater garbage target detection algorithm based on improved YOLOv7-Tiny

In the field of object detection, deep learning has been used extensively, especially in algorithms like Yolov7, which have achieved significant accuracy improvements. However, traditional convolutional neural networks are computationally intensive and require powerful GPU support, making its deployment on embedded devices difficult. This presents a problem for researchers as the high device requirements hinder their related research work. Consequently, more people opt to use lightweight networks, such as Yolov7-tiny. However, when using Yolov7-tiny for underwater garbage detection, it has been observed that while achieving good accuracy in mAP (mean Average Precision) at IOU (Intersection over Union) of 0.5, the performance is not satisfactory in the mAP range of 0.5 to 0.95. This limitation may be attributed to the trade-off in network performance during the process of model lightweighting. To address these issues, An enhanced Yolov7-tiny method for the detection of underwater trash objects is proposed in this paper. First, the algorithm employs an enhanced Ghost convolutional feature extraction module, which starts with conventional convolutions using a smaller number of channels, then performs grouped convolutions to obtain partial output maps with features. Finally, the maps with features obtained from the first convolutional step are added to the channels obtained from the second grouped convolution step. This design effectively reduces model complexity while extracting richer feature information. Secondly, the algorithm utilizes the CA (Channel Attention) mechanism to weight channels based on their positional information, thereby efficiently extracting features. The network can concentrate more on important feature regions by learning position weights in the feature maps. Lastly, the algorithm combines the Repeated Weighted Bi-directional Feature Pyramid Network (BIFPN) for feature fusion. BIFPN employs multiple down-sampling steps for short skip connections, enhancing the long-range correlations of encoded features. The results indicate that the object detection algorithm performs better and multi-level feature information may be integrated more effectively. Through experiments on public datasets, the proposed algorithm demonstrates superior accuracy, speed, and computational effectiveness in contrast to the Yolov7-tiny network's initial design. The improvements and combination of these modules effectively enhance feature representation, extract essential information, and boost detection accuracy, thereby elevating the overall performance of the object detection algorithm.