In order to achieve real-time underwater object detection, many researchers have chosen the YOLO series as the basis for further development. Wang et al. [21] reduced the network model size of YOLOv3 and replaced batch normalization with instance normalization in some early layers, thus enabling underwater deployment while improving detection accuracy. Al Muksit et al. [22] developed the YOLO-Fish network by modifying the upsampling steps and adding a spatial pyramid pool in YOLOv3 to reduce the false detection of small fishes and to increase detection ability in realistic environments. Zhao et al. [23] and Hu et al. [24] improved the detection accuracy by optimizing the network connection structure of YOLOv4 and updating the original backbone network.

YOLOv5, like other YOLO series, is a one-stage object detection algorithm. YOLOv5 includes four variants, namely YOLOv5s, YOLOv5m, YOLOv5l, and YOLOv5x, whose network size and number of parameters increase successively. Figure 1 shows the overall structure of the YOLOv5 network. An image is first processed by the backbone for feature extraction, followed by the neck for feature fusion, and finally is outputted as the head for the prediction of objects. The backbone is used to extract features from images, containing a total of 53 convolutional layers for various sizes of image features. The CSPDarknet53 structure proposed in YOLOv4 is slightly modified and continuously employed in YOLOv5 [25]. The spatial pyramid pooling module in YOLOv4 is replaced by the spatial pyramid pooling-fast module to improve computational efficiency. The neck is used to reprocess the extracted features for various spatial scales, which consists of top-down and bottom-up paths to form a cross-stage partial path aggregation network [26]. The feature pyramid network [27] is used to fuse the features from the top to the bottom, and a bottom-up path augmentation is used to shorten the path of low-level features. The fused features are integrated into the head structure, and three prediction paths are used. Each path fuses the low-level features of different receptive fields and finally outputs the bounding box, confidence, and category of the detected objects.

Mosaic, mix-up, copy–paste, and several other methods are employed in YOLOv5 for data augmentation [10]. YOLOv5 uses complete intersection over union (IoU) loss [28] to compute the bounding box regression loss, in which the distance between the center points of the bounding box and ground truth (GT) and the aspect ratio of the bounding box are rigorously considered. Similar to YOLOv4, the basic anchor size is computed in YOLOv5 using the k-mean algorithm, but YOLOv5 introduces a priori judgment called auto-anchor to enhance the versatility of anchor boxes. If the preset anchor size matches well with a dataset, the recalculation of anchor boxes can be avoided. YOLOv5 inherits the non-maximum suppression method from its predecessors [29], in which two bounding boxes are considered to belong to the same object if the IoU of two bounding boxes is higher than a certain threshold.

The proposed TC-YOLO network was developed based on YOLOv5s. The overall structure of YOLOv5 and the data augmentation methods were preserved, while self-attention and coordinate attention mechanisms were integrated with the backbone and the neck, respectively.

Most of the targets for underwater detection are small and dense, so researchers have introduced attention mechanisms to improve the detection performance. Sun et al. [30] attempted to design a new network using Swin Transformer as the backbone for underwater object detection and obtained similar performance as the Cascade R-CNN with the ResNeXt101 backbone. Other researchers chose to combine attention mechanisms with existing networks to improve the detection accuracy for underwater targets [31,32]. For example, Wei et al. [20] added squeeze-and-excitation modules after the deep convolution layer in the YOLOv3 model to learn the relationship between channels and enhance the semantic information of deep features.

The attention mechanisms used in computer vision are traditionally divided into three types: spatial attention, channel attention, and hybrid attention. Applying the attention mechanism in the spatial domain gives neural networks the ability to actively transform spatial feature maps, for example, the spatial transformer networks [33]. Applying the attention mechanism in the channel domain enables strengthening or suppressing the importance of a channel by changing the weight of this channel, for example, the squeeze-and-excitation networks [34]. Later works, such as the convolutional block attention module (CBAM), employ a hybrid mechanism that combines spatial attention and channel attention modules together and is widely used in convolutional network architectures [35].

Coordinate attention (CA) is another hybrid attention mechanism recently proposed in 2021 [36]. Improved from channel attention, CA factorizes channel attention into two 1-dimensional feature-encoding processes, each of which aggregates features along one of the two spatial coordinates. It encodes both channel correlations and long-range dependencies with precise positional information in two steps: coordinate information embedding and coordinate attention generation. Compared with the CBAM, which computes spatial attention using convolutions, the CA can preserve long-range dependencies that are critical to vision tasks. Additionally, the CA avoids expansive convolution computations, improving its efficiency compared with other hybrid attention mechanisms and enabling its application to mobile networks.

In addition to the above attention mechanisms commonly used in computer vision, Transformer was introduced in 2017 for natural language processing [37], and since then, it has been successfully applied in different neural network architectures for various tasks. Convolutional networks have the problem of limited perceptual fields. Multi-layer stacking is required to obtain global information, but as the number of layers increases, the amount of information is reduced, such that the attention of extracted features is concentrated in certain regions. Transformer, on the other hand, employs the self-attention mechanism that can effectively obtain global information. Furthermore, the multi-head structure used in the Transformer allows better fusion and more expressive capability, in which feature maps can be fused in multiple spatial scales. The Transformer mechanism has been applied to computer vision, such as Vision Transformer [38] and Swin Transformer [39]. However, Transformer modules require a large amount of computation, so they are mostly adopted in large networks and hardly used in mobile networks.

Label assignment aims to determine positive and negative samples for object detection. Unlike labeling for image classification, label assignment in object detection is not well defined due to the variation in bounding boxes. Zheng et al. [40] proposed a loss-aware label assignment for a one-stage detector for dense pedestrian detection. Xu et al. [41] proposed a Gaussian receptive field-based label assignment strategy to replace IoU-based or center-sampling methods for small object detection. Optimal transport assignment (OTA) was proposed in 2021 for object detection [42]. OTA is an optimization-theory-based assignment strategy, in which each GT is considered a label supplier, and its anchors are regarded as label demanders. The transportation cost between each supplier–demander pair is defined by the weighted summation of its classification and regression losses, and then the label assignment is formulated into an optimal transport problem, which aims to transport labels from GT to anchors at minimal transportation cost.

In YOLOv5, the ratios in height and width between a GT and an anchor box are computed as 𝑟_ℎ and 𝑟_𝑤, and then max(𝑟_ℎ,𝑟_𝑤,1/𝑟_ℎ,1/𝑟_𝑤) is compared with a threshold value of 4. If the maximum is smaller than 4, a positive sample is built for this GT; otherwise, the sample is assigned as negative. This label assignment mainly considers the difference in the aspect ratio between the GT and its anchor boxes, which is a static assignment strategy. Improved from the earlier versions of YOLO, a GT can be assigned to multiple anchors in different detection layers in YOLOv5. However, YOLOv5’s label assignment is still not sufficient for underwater object detection because the same anchor may be assigned multiple times, resulting in missed detection.

The quality of underwater images is often poor due to harsh underwater conditions, such as light scattering and absorption, water impurities, artificial illumination, etc. Enhancement and restoration techniques are applied to underwater images to improve the performance of underwater object detection. Li et al. [43] proposed an underwater image enhancement framework consisting of an adaptive color restoration module and a haze-line-based dehazing module, which can restore color and remove haze simultaneously. Li et al. [44] developed a systematic underwater image enhancement method, including an underwater image dehazing algorithm based on the principle of minimum information loss and a contrast enhancement algorithm based on the histogram distribution prior. Han et al. [45] combined the max-RGB method and the shades-of-gray method to achieve underwater image enhancement and then proposed a CNN network to deal with low-light conditions. There are other methods used for underwater image enhancement, including spatial domain methods, transform domain methods, and deep learning methods, among which spatial domain methods are considered the most effective and computationally efficient [46].

Several spatial domain image enhancement methods are introduced below. The histogram equalization (HE) algorithm computes the grayscale distribution of an image, stretches its histogram evenly across the image, and therefore improves the image contrast for better separation between foreground and background [47]. However, the HE algorithm increases the sparsity of grayscale distribution and may cause a loss of detailed image-related information. Adaptive histogram equalization (AHE) improves the local contrast and enhances edge details by redistributing the local grayscale multiple times, but it has the disadvantage of amplifying image noise [48]. On the basis of AHE, contrast-limited adaptive histogram equalization (CLAHE) imposes constraints on the local contrast of the image to avoid excessively amplifying image noise in the process of enhancing contrast [49]. Computations of field histograms and the corresponding transformation functions for each pixel are very expensive. Therefore, CLAHE employs an interpolation scheme to improve efficiency at the expense of a slight loss of enhancement quality. Histogram equalization methods only solve the problem of image brightness without adjusting image colors. The Retinex theory can be applied for image enhancement to achieve a balance in three aspects: dynamic range compression, edge enhancement, and color constancy. There are several image enhancement methods developed based on the Retinex theory [50], such as single-scale Retinex, multi-scale Retinex, multi-scale Retinex with color restoration (MSRCR), etc. Retinex-based methods can perform adaptive enhancement for various types of images, including underwater images, but the computational cost is much higher than histogram equalization methods.