Ensuring the quality and safety of iced aquatic products is the key to improve the benefits of the aquatic industry. Traditional aquatic product freshness evaluation faces the following challenges: complicated operations, samples destruction and low efficiency. An effective and scientific method is urgently needed for aquatic products cold chain system. To solve the above problems, an improved object detection network SSD was proposed to realize the sensitive area location and pomfret freshness evaluation. Firstly, image datasets of iced pomfret freshness grade was established based on environmental factors-pomfret image-total volatile basic nitrogen (TVB-N). The image data of pomfret was collected according to the physicochemical index TVB-N of pomfret freshness at a constant temperature of 0℃, with days as the unit of time. Then, the samples of pomfret images were expanded with data augmentation and marked by LabelImg, annotated image datasets of iced pomfret were provided for freshness detection. Secondly, based on prior knowledge, the eyes and gills of the pomfret were chosen as region of interest. Considering the trade-off between detection accuracy and speed in cold chain application, one-stage object detection network SSD performed better. SSD significantly improved the performance by replacing the backbone network and designing adaptive prior boxes. The improved SSD reached mean average precision of 98.97% and 99.42% on the golden and silver pomfret datasets respectively and the detection speed reached 37 frames per second. The results met the demand for real-time and assessment accuracy in application scenarios, and enabled low-cost, efficient and accurate assessment of pomfret freshness.